新着違うモデルを出品する販売者をどう見抜き、どう返金するか。

このページはまだ翻訳されていないため、英語で表示しています。

Changelog

Fees today: no MartLLM fee on requests. Adding credit by card costs 5.5% + $0.50, added on top. Fees named in older entries below (a 10% request fee, a 5% credit fee) are history: the 10% fee ended on 2026-10-22 and the 5% fee on 2026-10-29.

2026-11-20 — UberLLM is now MartLLM

  • New name, same service. UberLLM is now called MartLLM. Your account, credit, API keys (they still start with ull_) and request history are unchanged, and the price of every request stays the same.
  • Nothing to change in your code. The API address you use today keeps working for every request, and so do old links to our pages. Response headers keep their names (X-UberLLM-Cost-USD and the others), and media callbacks are still signed with UberLLM-Signature.
  • Entries below this one are from before the rename and keep the old name.

2026-11-20 — A simpler app: one menu, clean lists, Logs

  • New layout. One menu on the left (Models Market, API Keys, Logs, Billing, Docs, Sell capacity) with one Add credits button, and your account at the bottom. Affiliate Program is in the account menu. On phones: a top bar and a tab bar at the bottom.
  • Logs (was API Usage under your keys): every request, what it cost and what you saved, with quick filters. What you send and get back is deleted after 90 days; a key can be set to save none of it. Old links still open it.
  • Models Market starts with Seedance 2.5, DeepSeek and the newest frontier models. Each price is the cheapest offer from the sellers on the market, with how many offers there are.
  • Billing uses the same add-credit box as Home. Past payments with receipts underneath.
  • Cleaner lists everywhere, and each screen keeps only what it is for: keys on API Keys, requests on Logs, money on Billing.

2026-11-20 — Analytics: where your money goes, what you saved, speed and failures

  • New screen: Dashboard → Analytics. For the last 24 hours, 7, 30 or 90 days, next to the period before it: what you spent and how it moved, what you saved against official prices, and how many days your balance lasts at that pace.
  • Where the money goes: spend per day split by model, key, type (text, images, videos, audio) or one of your own tags. Send "metadata": {"client": "acme"} with a request and you can see, and download, what each client, campaign or user cost. Click a row to see only that.
  • Images, videos and audio: how many you made per model and size, the cost of each one, the cost per second and the typical wait.
  • Reliability and speed: requests that worked, failed or were stopped early, the most common reasons in plain words with what to do about each, and the typical wait for the first word and the whole answer over time and per model.
  • Tokens: what you sent, how much was read from cache, and what was written back.
  • Built for developers. Errors are split into what you fix in your request or key and what happened on our side, with the HTTP status and error code your app got. Also: writing speed in tokens per second, slowest 1% waits, your 20 slowest requests, the busiest minute, a weekday × hour traffic grid, the share sent with streaming, requests by size, and all models side by side.
  • The same numbers as JSON with your API key: GET https://api.martllm.com/v1/analytics. See Analytics.

2026-11-20 — Home page models: Seedance 2.5, DeepSeek and the frontier models

  • The home page's model cards are now Seedance 2.5, DeepSeek V4.1 Flash and the newest frontier models: Claude Opus 5.5, GPT-6.2 Sol, Gemini 4 Hydrogen, Claude Fable 5.2, Grok 4.7 and Kimi K3.1 Pro, each with its live price. Seedance 2.5 is the video model customers spend the most on here, and DeepSeek V4.1 Flash is the most used model by far.
  • The heading no longer says "Top models, live on the market": it is a hand-picked list, so it now reads "Models you can call now".

2026-11-20 — A receipt for every payment, and API Usage in one place

  • Every paid top-up now has a receipt, whatever paid it (card, Apple Pay, Google Pay, PayPal and the rest). Open Billing → Your payments → Receipt: it shows the receipt number, date, your account, our company details, the credit, the payment fee and the total, and prints or saves as a PDF. Before, only the few payments made through Stripe had one.
  • API Usage now sits directly under your keys on Dashboard → API Keys. The old address and the API Usage button on Billing open it there.
  • The dashboard home no longer shows a "Tokens: 0" tile for image, video and music jobs, which have no tokens.
  • The Models Market says "Updated 11 h ago" instead of "Live · updated 690m ago" when its prices are an hour or more old.
  • Your account name shows at once in every new tab.

2026-11-20 — "How we protect your data" next to "How we check sellers"

  • The model checks page now says how we protect what you send, in all 12 languages (martllm.com/verification#data): API keys, passwords, tokens and private keys are swapped for tags before any seller sees a request, for every key; email addresses, phone numbers, card numbers and IP addresses too with the key setting "Hide personal data from providers" (on for new keys); requests are never used to train models; and the X-UberLLM-Hidden header shows what was hidden. It also says plainly what is not hidden (names, street addresses, images and files) and links to Privacy: what the provider sees.
  • The add-credit card in the dashboard has a matching line under "How we check sellers": passwords and API keys in your requests are hidden from sellers, with a link to the details.
  • Nothing changes in how requests are handled: this explains protection that is already on.

2026-11-19 — Several keys per direct seller, kept in our database

  • We can now hold more than one account key at each direct image and video seller. When a seller says one of our keys is sending too many requests, has run out of balance or is refused, that one key is rested and your job is tried again on our next key at the same seller before it moves to a dearer seller. Before, one busy or empty key took the whole seller out for every job.
  • A key that is rate limited comes back by itself after 15 minutes. A key that is refused or out of balance stays off until our team has checked it. The seller itself stays on while it has another working key.
  • A job is followed with the key that started it, so a running job is not lost when that key is rested or replaced while it runs.
  • Nothing changes in your requests, prices or bills. This is on our side only: the keys are sealed in our database instead of server settings, and no key is ever shown again after it is saved (only its last 4 letters).

2026-11-19 — Listed image and video prices are the price you are quoted

  • The price a list shows for an image or video model is now worked out the same way as the quote for your request. Before, the lists (model pages, GET /v1/models, the public catalog) showed the seller market's price, while a request was priced from the direct sellers too. Nano Banana Pro edit was listed at $0.1455 a picture and billed about $0.02. The two now come from one piece of code, so they agree.
  • Image models priced by size now list each size. A model that costs more at 4K than at 1K or 2K has one row per size in GET /v1/models (pricing.resolutions) and in the public catalog (media_by_resolution, with usd and list_usd per picture). Nano Banana Pro today: 1K and 2K $0.01995, 4K $0.029925.
  • A video a seller prices by the whole clip is listed per second at its dearest rate, which is its shortest clip, and the row says which length that is (seconds). Gemini Omni Flash 1.1 at 720p: a 4-second clip costs $0.249375, so the list says $0.062344 per second for 4 seconds; a 10-second clip costs $0.399, less than 10 listed seconds. No clip length costs more than the listed seconds.
  • A model only a direct seller makes right now is listed as available, at that seller's price. Before, it read as unavailable while requests for it worked.
  • The lists follow our sellers within about 30 seconds. When a direct seller is paused or switched off, or its price or our percent on it changes, the listed price moves with the quote. Where a list showed the price of a seller that is switched off, it now shows the price requests are really billed, which can be higher.
  • The listed price is the price a key set to Balanced (the default) is quoted. A key set to Lowest price pays that or less. A key set to Fastest can pay more when a faster seller takes the job, never more than the model's list price (list_usd) for the request.
  • A listed price per second is rounded to the nearest $0.000001, so listed seconds times your clip length can differ from the quote by a few millionths of a dollar. A Fal video keeps its own price table (media_rates), as before.

2026-11-19 — The default setting keeps the cheapest seller for images and video

  • A key that never changed its routing setting keeps the cheapest seller. The default now steps over a seller only when it keeps failing or is more than twice as slow as the fastest one we have timed. It no longer moves a job to a dearer seller for a 20% gain in speed.
  • Balanced is now its own choice ("routing": "balanced" on PATCH /api/v1/keys/{id}): the cheapest seller that is at most 20% slower than the fastest. Pick it when a faster result is worth a higher price.
  • Failures count. A seller that made fewer than 7 in 10 of its recent jobs goes after the healthy sellers on every setting, and a seller's speed is compared as the time a job is expected to take there, lost jobs included.
  • One seller, several speeds. A seller can now offer a model in more than one way (a cheaper, slower one and a dearer, faster one). Each is timed by itself and your key's setting picks between them. Gemini Omni Flash gets two new ones today.
  • The seller market is ranked with the direct sellers, at its own price and measured speed, so Fastest and Balanced use it first when it is the quicker one. If it cannot take the job, the next seller in your key's order does.

2026-11-19 — Balanced, Lowest price and Fastest now work for images and video

  • Your key's routing setting now chooses the seller for image and video requests too. Until today it only applied to text models, so three keys set to Balanced, Lowest price and Fastest all sent the same video to the same seller at the same price.
  • How each setting picks, when more than one seller makes exactly what you asked: Lowest price takes the cheapest. Fastest takes the seller we have timed as quickest; it may cost more than the cheapest, but never more than the model's listed price for your request. Balanced (the default) takes the cheapest seller that is not more than 20% slower than the quickest one.
  • Speed is measured, not guessed: the usual time from submit to done over each seller's recent finished jobs for that model. A seller is compared on speed once it has finished 3 jobs in the last 7 days; until then every setting picks the cheapest, as before.
  • The price you are quoted is the price you are billed. A quote ("validate": true) and a job's estimated_cost_usd show the price of the seller your setting picks. The price lists show the price a Balanced key is quoted (see the entry above), so a Fastest key can pay more than the list when a faster seller takes the job; set the key to Lowest price to always get the cheapest. Details: Your key's routing setting.
  • Some image and video prices can now go lower by themselves. A seller can be set to price its jobs 1% under the cheapest other marketplace we track for the same model, when its usual price is higher. A job is billed the price it showed at its start.

2026-11-19 — Markup per supplier

  • Our margin on a direct video or image seller is now set per seller. Until today every job a direct seller made was billed that seller's charge plus a fixed 5%. Each seller now has its own percent, and all start at 5%, so no price changed today.
  • The quote, the price a job shows, the hold and the bill use the same percent, and so does the price of each resolution on a model page and in GET /v1/models (pricing.resolutions). When several direct sellers make the same request, it goes to the one that costs you least.
  • A job keeps the percent it started with. If a seller's percent changes while your job runs, the job is billed the price it showed at its start.
  • Jobs made on the seller market are not affected.

2026-11-19 — Video prices by resolution

  • A video priced by resolution now shows the price of each resolution, on its model page ("Price by resolution", per second and for a 5-second clip) and in GET /v1/models (pricing.resolutions). Before, the lists showed one number, the cheapest resolution's: Seedance 2.5 read $0.107 per second while a 720p clip is billed $0.227 per second.
  • Each resolution's price is the one a clip is billed at: the cheapest seller that makes it at that resolution, with our 5% in. Seedance 2.5 today: 480p $0.10, 720p $0.227, 1080p $0.56 per second.
  • A Seedance 2.5 clip made on the seller market is never billed above the direct seller's price it showed. The market gets these clips only when one of its sellers is cheaper, and on November 19 its failover between sellers made 9 clips cost $1.18-1.21 against the $1.13 shown. Those clips now bill at most the price shown.

2026-11-19 — Cheapest GPT, Gemini, Grok, Kimi, GLM and Qwen API: dated comparisons

  • Six new comparison pages, like the Claude and DeepSeek ones: GPT, Gemini, Grok, Kimi, GLM and Qwen. Each sets UberLLM's price now beside the maker's own price and the prices of OpenRouter, Vercel AI Gateway and Requesty (plus DeepInfra and Novita for open models), model by model, and says where someone else is lower.
  • The maker's prices were read on its own pricing page on 2026-11-19, with the link on the page. A price change the maker has already announced is applied on its date: Gemini 4 Argon and Carbon go to $4 / $20 per 1M tokens on December 1, 2026.
  • A page is built only while UberLLM, with its card fee added, is below the maker on most of the models compared, and it is rebuilt every hour from the live prices.

2026-11-19 — Secrets and personal data hidden from the seller

  • Secrets are hidden for every key. Before a chat, messages or responses request leaves UberLLM, API keys, passwords, tokens, private keys and passwords inside links that we recognise are swapped for tags like [SECRET_3f9a2c1b]. The real values are put back in the answer, streamed answers and tool-call arguments before you get them.
  • New key setting: Hide personal data from providers. It also swaps email addresses, phone numbers, card numbers and public IP addresses. It is on for new keys; turn it on for older keys in Dashboard → API keys.
  • See what was hidden in the X-UberLLM-Hidden response header (for example email=2,secret=1). Names, street addresses, free text and files are not hidden. Details, and when to turn the setting off: Privacy: what the provider sees.

2026-11-19 — Ad creative recipes: from one product photo to image and video ads

  • Four tested recipes for teams that make ad creatives. A person holding your product with a 9:16 clip and a voice-over line, static variations from a list, one still and one clip per row of a CSV or Google Sheet on a schedule, and a studio spin video. Each is an n8n workflow to import by link and one command of a script with no dependencies, in the templates repository. What each made, how long it took and what it was billed on a real run is on the ad creative page.
  • A skill for coding agents. npx skills add PSA-Team-source/uberllm-n8n-templates --skill ad-creative gives Claude Code and other agents the same recipes. It prices a run first and asks before it spends.
  • New guide: Ad creative workflows. How to start a job, wait or take a callback, download a file before its link ends, and keep the cost under a ceiling.
  • The UGC product ad n8n template fixed. It handed the edited picture's own link to the video model, which cannot open a link that needs your key, so the clip could fail (at no charge). It now uploads the picture as a file first, and so does the faceless Shorts pipeline. Both were run again on 2026-11-19.
  • Sitemaps. A translated guide that is still shown in English (its translation is not ready) is no longer listed in that language's sitemap.

2026-11-19 — Video and image jobs: the price you are shown, captions, clearer refusals

  • A video or music job shows the seller's real price from the start. estimated_cost_usd on a job is now the figure of the seller that took it whenever that is above the listed price, so what you see when the job starts is what you should expect to pay. Before, a job could show the listed price and be billed more (a 5-second Kling v3 Standard clip showed $0.39 and cost $0.588). max_cost_usd is, as before, the most a job is ever billed. Quotes (validate and POST /v1/quote) now say this in their note: cost_usd is the listed price, the bill is what the provider charges, never above max_cost_usd.
  • max_cost_usd from the start reply always holds. A job whose first seller lost it could be moved to a dearer seller and billed above the max_cost_usd it started with (a Kling O3 clip that started at $0.3675 was billed $0.4531). A job is no longer moved to a seller that would cost more than its max_cost_usd. And estimated_cost_usd now also rises when the seller's figure arrives a few seconds after the start, not only in the first reply.
  • Captions. /v1/audio/transcriptions now answers response_format=srt and vtt with a real caption file with timings, and verbose_json with timed segments. These were listed before but refused. See the media guide.
  • aspect_ratio on image-to-video. A model that makes the clip in the shape of your source picture (its schema lists no aspect ratios) now answers a shape with a clear 400 before anything is charged, instead of a failed job. Send a picture that already has the shape you want.
  • Slow pictures get more time. A picture made by one of our direct sellers now gets 88 seconds instead of 75 before the request gives up (about one in eight of those pictures was cut off).
  • Clearer answers when a job cannot be made. A prompt or picture a seller turns down under its content rules now says so (it read "the video could not be made" and invited a retry). A model its seller stopped making answers "no one is serving it" and moves on to your backup models, instead of "Request body must include a valid 'model' field". A seller that only takes pictures as data is now sent your linked picture inline, so the edit goes through.

2026-11-19 — Lighter pages

  • Model comparisons are listed for search engines in English only. The same comparison in the 11 other languages still opens from the language links, but search engines are asked not to list those copies, so they spend their visits on model, price and answer pages.
  • Two guides you can copy from: Claude Code on DeepSeek, GLM, Kimi or Qwen now has the settings for each model family, what changes on a model that is not Claude, the errors people hit and how to switch back; and a new page answers is there an LLM API marketplace where sellers compete on price?. Also: every page now counts uncensored models the same way (only those a seller is offering right now), and the pricing page says credit can be paid by card or PayPal.
  • The home page and the model list are about a sixth of their old size. They now carry the top models and read the rest of the list after the page opens, so both open faster and search engines can read them whole. The sitemap is split by page type, and old model links that had stopped working lead to the model's current page again.

2026-11-18 — Fewer failed calls: listed prices you can really buy, and a clear answer for prompts with secrets

  • A key set to "never pay more than the listed price" now keeps working when sellers get pricier. When no seller is as cheap as the listed price, the listed price moves up to the cheapest price sellers really fill, within minutes, for the next two hours. Before, it followed the market only hour by hour, and such keys could be refused for a few hours.
  • Prompts that contain a secret. API keys, passwords, private keys and tokens in a prompt are hidden from the provider. When one cannot be hidden, the request is tried again, and if it still cannot be sent you get a clear answer (400 secret_in_request: remove the secret and send again) instead of "no one could serve, try again". You are never charged for it.
  • Kimi K3 with a temperature or top_p. Kimi K3 and K3.1 Pro run at a fixed temperature and top_p, and refused any request that set its own ("Parameter 'temperature'=0.5 is not supported"). Such a request is now sent again without that setting and answered at the model's own value. A value out of range is still refused, as before.
  • Model comparison pages in search: which "X vs Y" pages search engines are sent now follows each model's use over the last 7 days, not the last 24 hours, so the list stays the same from one hour to the next. The public models feed has a new activity_7d_requests number per model.
  • Search results: every model-comparison page now has its own description naming both models, and no page description runs past 160 characters.

2026-11-18 — Free uncensored chat, and answers to the most-asked questions

  • Uncensored AI chat: chat with an uncensored model in the browser with no account, 12 messages a day. It is for adults, and two rules stay: no sexual content involving minors, and no help with breaking the law or harming people. With an account the same models have no daily message limit and work through the API.
  • Answers has 33 new pages for questions people ask most: how to get a Claude, DeepSeek or Kimi API key; what the OpenAI, Claude, Gemini, DeepSeek, Grok and Kimi APIs cost (list price next to the price billed now); whether Claude Code is free; OpenRouter's free-model limits; paying by PayPal; hard spending limits; and whether prompts are stored.
  • Facts about other services on those pages were read on their own sites on 2026-11-18 and are linked under each table.
  • New models: every model added in the last 30 days that takes requests now, newest first, with the price billed now next to the official price. Rebuilt every hour.
  • Set-up for every tool, on every chat model's page. Under "Use this model in your apps", a model page now gives that model's own settings for Claude Code, Codex CLI, OpenCode, Cursor, Cline, SillyTavern, Janitor AI and Open WebUI: the base URL, the exact model id, the context numbers the tool needs, and which address the tool calls (OpenAI chat, Responses or Anthropic Messages). Coding tools are listed only for models that take tool calls. Each row on New models links to it.
  • A page for one model in one tool. For the newest models and the most used ones (40 models), each tool now has its own page with that model's settings, for example Kimi K3.1 Pro in Claude Code: the block to copy, the model id, the three API addresses, the context size the tool needs and the price billed now. The model page and New models link to these pages, in all 12 languages.
  • The model checks page now says in words what it means when no directly listed seller offer is taking requests, in place of a bare zero.
  • About names the founder: Alex, Founder, reachable at [email protected]. The SillyTavern and Janitor AI set-up guides list only models that are in the catalog now.

2026-11-13 — Dashboard: the numbers behind adding credit

  • The Add credit card on a new account's dashboard, and the Billing page until your first payment, now show live numbers: paid requests served in the last 24 hours, what buyers saved against list price in the last 30 days, test requests sent to sellers in the last 7 days, and how many seller offers failed our tests and were stopped.
  • Under them: what protects your payment (credit never expires, unused credit refundable for 30 days, a failed request is not billed, no UberLLM fee on requests), why prices are below list, and the company that runs UberLLM.
  • The notice about how we catch a seller listing the wrong model now also shows at the top of every dashboard screen. Close it and it stays closed.

2026-11-13 — Sellers: list unused provider credit from the terminal

  • npx uberllm sell checks your provider key with the provider, shows the remaining balance where the provider has an API for it (OpenRouter, DeepSeek, Venice AI, Moonshot AI / Kimi, Ollama Cloud, Morpheus), and lists the models you choose only after you type yes. The key is sent to the provider, and to UberLLM only after your yes. See Sellers: from the terminal.
  • Guides by provider. Sell unused AI credits and one guide for each verified provider, 23 in all. Each says what the provider's terms said about reselling when we read them. Many do not allow it without the provider's permission, and /sell now says so too.
  • The seller form opens with a provider chosen from a link: /seller/new?provider=<id>.

2026-11-13 — Embed live prices on your own site

  • A price table for your site. Paste one script tag, or an iframe where scripts are not allowed, and your page shows UberLLM's current prices for the models you pick. The table updates by itself and needs no API key. Choose the models, the type (text, image, video, audio), light or dark, and one of 12 languages.
  • The pricing pages have an Embed a live price table link and each model page has Embed this price, with the code ready to paste. Model pages also give the README price badge. Embed live prices on your site.

2026-11-13 — One command to set up your coding tools

  • npx uberllm init finds Claude Code, Codex CLI, OpenCode, Kilo Code, Goose, Aider and Continue on your machine and sets each one up to use UberLLM. You connect your account once in the browser; the key is saved to a file only you can read.
  • Every file is backed up before it is changed. --dry-run shows the changes first, --project sets up one project only, and --undo puts everything back.
  • For Cursor, Cline and Roo Code, which keep these settings inside the editor, it prints the three steps to follow. uberllm init.

2026-11-13 — Price API: current prices as JSON, no key needed

  • GET /api/v1/public/prices returns the current price of every model you can call, as JSON, with no API key. Each model has its price now, its list price and how far below the list price it is. These are the same numbers as on the model's page. Filter with model, type, q and limit. Any website can read it from the browser.
  • /api/v1/public/badge/<model-id>.svg is the same price as a small picture for a README or a web page. It updates by itself.
  • The JSON address allows 300 requests a minute per IP address. Badges have no request limit. Price API and price badge.

2026-11-13 — Starter kits and a live price badge for your README

  • Starter kits. Four small projects to copy and run on UberLLM: a Next.js chat app on the Vercel AI SDK, a LangGraph agent, a CrewAI crew, and a tool-calling agent CLI in one TypeScript file with no dependencies. Each pins its dependency versions and has a smoke test. Copy one with npx degit PSA-Team-source/uberllm/starters/<name>.
  • Price badge. If your project supports UberLLM, a badge in your README shows the current price of the model you recommend and updates by itself. The Markdown and HTML to paste are on the Starter kits page.

2026-11-13 — Share what you saved from Activity and Credits, and a public page

  • Share on more screens. The line "You saved $X vs list price in the last 30 days" and its Share button are now on Dashboard → Activity and Dashboard → Credits too (they were on Overview and Affiliate Program). One click makes your card. Nothing is shown until you have saved something.
  • A public page that stays up to date (in the Share window): one link that always shows your last 30 days. Off until you turn it on; it shows totals and model names only, never prompts, keys, your name or email. Details.

2026-11-12 — Do not save my requests and answers, per API key

  • Do not save my requests and answers (Dashboard → API keys, create or edit a key): UberLLM normally keeps each request and its answer for 90 days to fix failed requests and answer support questions. With this on, calls made with that key are not saved. We keep only the record of each call (time, model, token counts, cost, the address it came from, the providers we tried and any error they returned), not what you sent or what came back, so we cannot look into a wrong answer for you. Video and music jobs still keep their request, so a job a provider drops can be started again. Dashboard API: no_store on POST/PATCH /api/v1/keys. Details. The Privacy Policy says the same.

2026-11-12 — How we catch a seller listing the wrong model

  • New article: How UberLLM catches a seller listing the wrong model. It covers the four checks every seller offer gets (the model is matched at listing, tested before any buyer, approved by a person, and tested again on live requests), what happens to a seller caught faking, a real offer caught this week, and what no check can prove. Linked from the banner at the top of every page.

2026-11-10 — Minimum discount per API key

  • Minimum discount (Dashboard → API keys, create or edit a key): set a percentage, and that key's chat requests go only to providers at least that much below the model's official price. The bill never goes above that discount either. When no provider is that cheap at the moment, the request is refused and costs nothing. Empty means no minimum. Dashboard API: min_discount_pct on POST/PATCH /api/v1/keys. Details.

2026-11-10 — Choose how each key picks a provider; see the discount on every call

  • Routing setting per API key (Dashboard → API keys): Balanced (the default: fast and cheap; a faster provider priced above the listed price may answer), Lowest price (always the lowest-priced provider; never a faster, pricier one) or Fastest. Plus Never charge more than the listed price: a request is held to providers at or under the listed price when it came in and never billed above it, and is refused for free when none sells at that price. A request's own provider.sort or provider.max_price still wins. Details.
  • Activity shows the discount: for each chat call, what its tokens cost at the listed price at that moment and at the official price, the % below the official price you paid, prompt tokens read from cache, and how its provider was picked. A call billed above the listed price is marked and says why. The same figures are new columns in the Activity CSV (cached_tokens, listed_price_cost_usd, official_price_cost_usd, pct_off_official, routing, price_note).
  • Why: providers on the marketplace reprice minute to minute, and Balanced routing buys speed, so the same model could cost several times its listed price within minutes. You can now see that per call and turn it off per key.

2026-11-07 — martllm.com in Simplified Chinese (13 languages)

  • The whole site now reads in Simplified Chinese too, at /zh/ addresses (/zh/models/…, /zh/docs/quickstart, /zh/sell). That makes 13 languages: English, French, Spanish, German, Portuguese (Brazil), Italian, Dutch, Simplified Chinese, Vietnamese, Japanese, Korean and Indonesian.
  • Chinese-speaking browsers landing on an English page see a small banner offering the Chinese version. Clicking it sets a cookie and navigates to the /zh/ page; the banner stays dismissed for 90 days.
  • llms.txt has a Chinese section so AI assistants answering in Chinese can cite UberLLM with the right links and descriptions.
  • Chinese search engines (Baidu, Sogou, Shenma, 360) are welcomed in robots.txt, and /zh/ pages carry Chinese keywords meta and zh-Hans structured data for Baidu's indexer. New pages are pushed to Baidu's URL submission API on each deploy.

2026-11-06 — "Cheapest API" is now on each model's page

  • Each model page now shows the comparison that had its own page: UberLLM's price next to the same model's price at OpenRouter, Vercel AI Gateway, Requesty, DeepInfra, Novita AI (video: fal.ai and kie.ai), with the time the prices were read and a link to each list. It is shown only when UberLLM is lower on input and on output with the 5.5% card fee added. The old /answers/cheapest-<model>-api links go to the model page.
  • When the lowest offer on the market is below the listed price that GET /v1/models shows, the model page names both numbers.
  • Text model pages link the setup guides for Claude Code, Cursor, Cline, SillyTavern and Open WebUI (Janitor AI for uncensored models).
  • The home page lists the models used most on UberLLM in the last 24 hours, and each family page lists its most-used models first.
  • New page: Venice AI API alternative compares every text model both UberLLM and Venice AI sell, with Venice's live price list against UberLLM's price (card fee included). A model where Venice is cheaper says so. The page exists only while UberLLM is lower on most models (today 65 of 69). Available in all 13 languages.
  • Uncensored model prices now hold for 7 days: the price shown on the model page and in /uncensored-llm-api is the higher of the best seller's current price and the listed price over the past week, so the page never promises a price below what a request is billed. The SillyTavern and Janitor AI tables list every uncensored model that is taking requests.

2026-11-05 — Sellers list only from verified providers, and every offer is checked before buyers see it

  • Approved providers only. A seller can list only through a provider on our list, such as OpenRouter, Groq, Together AI, Fireworks AI, DeepSeek, Mistral AI, Z.ai, Alibaba DashScope, Meta or Amazon Bedrock, with its own key there. Your request runs on that provider's own servers, so the provider runs the model, not the seller. A seller's own server can no longer be listed. A seller can ask us to add a provider; we add it only after testing it.
  • Verified by default. Each provider is marked verified or not. Your requests reach only seller offers on verified providers, unless you turn on Also use sellers on providers we have not verified for a key.
  • The model must match. The provider's model must be in that provider's own model list and be the same model as the one listed, and the model named in every reply is compared with the model listed. An offer whose provider says its key is no longer valid is taken off automatically.
  • These come on top of the checks every offer already gets: an entrance test, a trial on copies of real requests, probation, a review by our team before any buyer, tests and a running score on live requests, and a ban with buyers refunded for a seller caught serving a different model.
  • New page: Model checks says what the checks show, what no software check can prove, and the live numbers.

2026-11-04 — OpenAPI spec

  • https://martllm.com/openapi.json is an OpenAPI 3.1 description of every customer route under /v1: chat completions, Responses, Anthropic messages, models, balance, images, video and music jobs, speech, transcription, quotes and files. Import it into Postman or Swagger UI, or generate a client from it.

2026-11-04 — Agent keys and Provider Market for sellers

  • Agent keys: a key that can call only the seller API (/v1/seller), so a script or agent can manage your offers without your login or your spending key. It can't make AI requests, spend credit or change where payouts go. See Agent keys.
  • Provider Market: each provider key's last 30 days: earned, calls served, win rate against all calls on the same models, and today's take against its daily limits (GET /seller/market).
  • Offers are grouped by provider (switch it off with Group by).

2026-11-04 — Seller console in tabs: Earnings, Payouts, Offers, New offer, Agent keys, Provider Market

  • Earnings: owed to you, paid out, requests and tokens served, and the average discount you gave, plus a day-by-day chart, today's numbers and demand by model. GET /seller/earnings now also returns daily and avg_discount_pct.
  • Payouts: the total not yet paid (available, on hold, on the way), where your payouts go, Pay me now, and your payout history.
  • Offers: health and active offers at the top, search, and filters for Live / In review / Needs attention. Saved provider keys are under Provider keys.

2026-11-04 — See the MCP tools before you sign up

  • The MCP server at https://api.martllm.com/mcp now answers initialize and tools/list without a key, so MCP directories, inspectors and your client can show its tools before you add a key.
  • Every tool call still needs your key. Without one it is answered 401 (with a WWW-Authenticate challenge) before any tool runs: nothing is read or spent.

2026-11-04 — Every seller offer is reviewed by a person

  • No buyer is sent to a seller's offer until a person on our team has approved it, on top of the automatic model check. New offers, and offers whose key, address or model changes, wait for review.
  • Sellers see such an offer as In review on the Selling page (listing_state: "review" and needs_review: true in GET /seller/offers), with its steps: Listed, Model check, Team review, Live. We email the seller when the offer is approved.
  • If our team does not approve an offer, it is paused and shows Not approved with the reason (review_note), and the seller is emailed the reason. Turning it on again after fixing it sends it back for review. See From listing to live.
  • Offers already listed were paused on 2026-11-04 and are reviewed the same way before they get buyers again. Buyers are not affected: every model keeps its regular route.

2026-11-04 — Every uncensored model in one list

  • Uncensored LLM APIs now lists every uncensored, abliterated and heretic text model on UberLLM with its price now (it showed 20 at most), and llms.txt has its own "Uncensored models" section with the same list.
  • Plain facts on both: no ID or KYC check (sign up with an email code or Google), prepaid credit by card or PayPal from $5, and the same request formats as every model: chat completions, the Responses API and Anthropic messages.
  • An uncensored model's page links straight to the list, and when it is offline it suggests other uncensored models first. A model page built while its only seller was offline now turns on in your browser as soon as the seller is back.

2026-11-02 — martllm.com in 8 languages

  • The whole site now reads in English, French, Spanish, German, Portuguese (Brazil), Italian, Dutch and Simplified Chinese. Public pages have their own address per language (/fr/sell, /de/docs/quickstart, /zh/models/…); English stays at the plain address. Search engines see every version of each page.
  • The dashboard, the seller console, sign-in and app connect screens follow your language: pick it from the language menu (sidebar, site menu or footer), or they use the language of the last page you opened, else your browser's.
  • Prices stay in US dollars, and model, provider and brand names stay as they are. The terms, privacy and refund policies are translated for reading; the English version is the one that applies.

2026-11-01 — Sellers state the precision of open-weight models

  • Open-weight models (GPT-OSS, Qwen, DeepSeek, GLM, Kimi, MiniMax…) can be run at full quality or compressed to 8-bit or 4-bit, which answers a little worse and can't be told apart by any test. A seller listing one now says how their server runs it: full quality (16-bit), 8-bit or below 8-bit.
  • Model pages show which models are open weights and how their sellers say they run them. In the API: open_weights on each model of GET /api/v1/public/models, and precision on each seller endpoint.
  • Offers listed before today show "Precision not stated" until the seller sets it. Stating a higher precision than the server runs gets the seller account suspended, with unpaid earnings used to refund buyers (Sell).

2026-11-01 — Key spend limits count what you really spent

  • A spend limit on an API key (daily, monthly, lifetime, or the daily cap) now counts what each request really cost. Before, it counted the most a request could cost (the amount held while it ran) and never corrected it, so a key could stop with "reached its spend limit" after a few cents of real use. Every key's count was reset to its real spend today.

2026-11-01 — Video prices say "per second", and Sora 2 is gone

  • Video prices now read the way they are billed: "$0.09 / sec" for a model billed per second of video (a 5-second clip costs 5 × that), and "per video" for a model with one flat price per clip. They used to read "/ job" for both.
  • Sora 2 and Sora 2 Pro are no longer listed. OpenAI shut Sora 2 down, and on 2026-11-01 no seller could take a request for either. Nothing was ever charged for those failed requests.
  • The "Official" price next to ours now comes from one source, the same on every page and in GET /v1/models: the model maker's own published price first, then OpenRouter's, then Fal's, shown only where it is above ours. About 100 more image, video and music models now have one, each kept with the page it came from and re-checked daily (Pricing).
  • Fal video prices follow the cheapest Fal seller again after a restart: a seller that ran one of our jobs in the last 24 hours stays priced in.
  • Every model shows its maker's logo in one standard style (models resold through a host show the real maker).

2026-11-01 — Sellers caught faking a model are banned, and buyers refunded

  • A seller caught serving a different model than listed is suspended: all its offers stop, its unpaid earnings are forfeited and used to refund the buyers it served, and it can't list again with the same key, server, payout address or email (Model check).
  • Seller earnings are held before they can be paid out: 14 days for new sellers (first 30 days, or under $500 earned), then 7 days. The Selling page shows Available to pay out and On hold; GET /seller/payouts adds held_usd and next_release_at.
  • Each offer now shows one plain state: Checking, Live, Paused or Suspended (listing_state in GET /seller/offers, replacing check_state and checked_at).

2026-11-01 — More video models show their official price

  • Video models whose seller market gives no price of its own now show OpenRouter's published price for the same model as the official price, on the models board, model pages and price answers. The detail view says "Official price · OpenRouter". The prices are read every hour and kept by UberLLM, so the API shows them too: media_ref on GET /api/v1/public/models (pricing).
  • It is OpenRouter's cheapest per-second rate for that model (lowest resolution, no sound), so the saving shown is never larger than it is. It is shown only where it is above our price. First models: MiniMax H3, Wan 3.0, Wan 3.0 Prime and Grok Imagine Video.

2026-11-01 — Affiliate Program: earn 15% of what the people you refer buy

  • The Share page is now the Affiliate Program. Creators with a real audience apply (where you post, links, a screenshot of your audience and how you'd share UberLLM); we review every application within 1-3 business days.
  • Approved affiliates earn 15% of all credit each account they referred buys in its first 30 days. Each commission is held 45 days and taken back if the purchase is refunded or charged back. A purchase paid with a card or PayPal account also used on the affiliate's own account is not counted.
  • Withdraw from $50, picking which commissions: UberLLM credit at once, PayPal or USDC on Base settled weekly and paid within 1-3 business days.
  • The page shows your link, earnings, every account you referred (emails partly hidden), every purchase with its status, and your withdrawals.
  • It replaces "give credit, get credit". No new invite gifts are offered; gifts already promised are still paid. All the rules: Affiliate Program.

2026-11-01 — Prices and discounts are the market's best, as the market shows them

  • Every price and "% off" on martllm.com is now the partner market's own "Best price" and "Best discount now", the same number its board shows, refreshed every 30 seconds. Before, pages showed the median price buyers paid over the last 24 hours, so a model could read 87% off here and 99.7% off there.
  • Discounts show one decimal, like the market (99.6%, 95%).
  • Summaries that said "X% less at the median" (against OpenRouter, Anthropic, DeepSeek list prices) now say "up to X% less".
  • What a request is billed has not changed: you pay what the seller that serves it charges, never more than the price cap on the model page.

2026-10-31 — A request a model can never take says so

  • Tools sent to a model that cannot call them, or a prompt longer than the model takes, used to end in "No one could serve this model right now, try again shortly": a retry that could never work.
  • On /v1/chat/completions (with no models list of your own) the closest model that can take the request now answers instead, and the reply's model names it.
  • When no model can, or on /v1/messages, you get 400 tools_not_supported or 400 context_length_exceeded with the reason in plain words and the models that can take it. Nothing is charged.

2026-10-31 — Add credit straight from your dashboard

  • A new account now adds credit on the dashboard itself: pick $10, $25, $50 or $100 and pay without leaving the page.
  • Next to the amount you see what it is worth at list price on the most-used models at today's best live market price: for example, $25 of DeepSeek V4 Pro calls would cost about $41,000 at list price (99.9% off). DeepSeek, Kimi, GLM, GPT and Claude are shown, plus how many Kling 3.0 videos the amount buys. The figures update every hour and are rounded down.
  • The same table is now on the home page, next to the headline. It opens with the biggest number ("$5 gets you up to $7,142 of AI at official prices"), then one line per model: official price struck through, you pay $5, and the % off to two decimals (DeepSeek V4 Pro −99.93% today). Kling 3.0 also shows how many videos $5 buys. $5 is now one of the amount buttons.
  • When your balance drops under $5, the dashboard shows one line with your balance and an Add credit button that opens the same form.
  • The OpenAI logo now shows on the dark pages (it was black on black).

2026-10-31 — Pay with PayPal when you add credit

  • The Credits page now has a PayPal button above the card form. Pay in PayPal's window and your credit is added the moment PayPal confirms, with no page reload. The fee is the same as by card (5.5% + $0.50, added on top).
  • Picking an amount opens the payment step in one click, and the card form shows up as soon as it is ready, with no loading screen.
  • A PayPal refund or dispute takes back the matching share of the credit, the same as for card payments.

2026-10-31 — Every seller offer is checked to serve the model it claims

  • We now check every seller offer: test requests that look like normal traffic, compared with the real model on token counts, hidden instructions, what the model says it is, and answers to test questions. Checks run soon after an offer is listed, then at random every few hours (Model check).
  • A new offer gets buyers once its first check passes, usually within about 15 minutes. An offer that fails is paused, and its seller's payouts are held while our team reviews it. The Selling page shows each offer's result in plain words; GET /seller/offers has it as check_state, check_reason and checked_at.
  • Padded token counts: when a seller reports far more prompt tokens than a request holds, the buyer is billed, and the seller paid, for at most about 2× our own count, plus a small allowance for tools and images.

2026-10-31 — Requests and answers are now saved for 90 days

  • Every API request and the answer we sent back are now saved for 90 days, then deleted. We use them to fix failed requests, answer your support questions (send us the X-Request-Id of a reply and we can see exactly what happened, including which providers were tried) and stop abuse.
  • Only our own team can read them. We never sell them or use them to train models. You can ask [email protected] to delete yours sooner.
  • Nothing changes in how you call the API. The Privacy Policy has the details.

2026-10-31 — 40 more video and image models, each at its cheapest price

  • New on Models: Seedance 2.5 (from $0.14 per second), Kling 4.0, Kling 3.0 Turbo, MiniMax H3, Hailuo 2.3, Hailuo 02 Pro, Vidu Q4, Wan 3.0 and Wan 3.0 Prime, PixVerse V6 and V6 Max, HappyHorse 1.1, Grok Imagine, Gemini Omni Flash, Seedance 1.0 Pro, Imagen 4 (Fast, Ultra), Ideogram V3, Seedream 3.0 / 4.0 / 5.0 Flash, GPT Image 1.5, Flux 3 Klein, Qwen Image 2.1 and Nano Banana. Call them like any other model: POST /v1/video/generations or POST /v1/images/generations (Media).
  • Video and image models we already listed now go to the cheapest seller serving them at that moment, and stay up when one seller drops out. The list shows that price, and you pay what the seller charged, never more than the quote.
  • Every video model is priced per second, and the price rises with resolution. Add "validate": true to see the exact cost for the length and resolution you pick before you spend (Check the price). Left out, the length is the model's shortest; the resolution sent is shown in the quote's defaults_applied and on GET /v1/models/<id>.
  • Requests a model can't make (a length or size it doesn't offer) are refused before anything is charged, with the lengths and sizes it does.

2026-10-31 — Invite friends: you both get credit when they add theirs

  • Your Share page can now pay an invite gift: when a friend signs up through your link (or your savings card) and makes a first top-up of the minimum or more, you both get the same credit. The page shows the amounts, for example "Give $5, get $5" when your friend adds $10 or more.
  • Nothing is given at sign-up: new accounts start at $0. A gift is taken back if the payment that brought it is refunded. One gift per new account; sellers' accounts don't qualify. All the rules: Invite friends.
  • The Share page shows Invited, Friends who paid, You earned and a list of every gift. Someone who arrives through your link sees their gift on the site and on the sign-up page, and again on their Credits page until their first top-up brings it.

2026-10-31 — Choose where you get paid

  • Sellers now choose where their payouts go: a PayPal email, or a USDC wallet on the Base network. Set it under Selling → Payouts → Where we send your money, or with PUT /seller/payout-method (Selling).
  • Our team sends each payout there, usually within 2 business days of it starting. A seller with money on the way and nowhere to send it gets one email asking where to send it.
  • One payout on the way at a time: what you earn meanwhile stays owed and goes into your next payout once the current one is paid.
  • We email you when your provider refuses your key or says the account has no credit, so you know why an offer stopped getting buyers (at most once a day).
  • A server's usage that cannot be right (a negative token count) is no longer billed; the reply's own text is counted instead. One seller lost $9.26 of earnings to this on 10-28/29; it has been paid back.
  • When your provider answers that your account is suspended or out of credit under any status code, buyers go to the next seller at once and your offer rests, the same as for a refused key.
  • Offers must use a public internet address. Local, private and UberLLM's own addresses cannot be listed.
  • Changing an offer's key (or its address) to one that passes the test puts it back in front of buyers at once, instead of after the rest from the old failures.
  • We email you when a payout is sent, with the amount, where it went and the payment reference.
  • No name to type before selling: Sell capacity goes straight to listing a model. You are shown to buyers as "Seller" plus 4 letters, never your email.

2026-10-31 — Sign in with Google in one click

  • Signed-out visitors see Google's "Continue as " prompt in the top right of martllm.com. One click signs you in, or creates your account if you are new. If you signed in with Google before, it signs you back in with no click at all.
  • If your browser hides that prompt, a "Sign in to UberLLM" card with the Google button shows once in the bottom right instead. The sign-in page keeps both Google and the emailed 6-digit code.

2026-10-30 — The company behind UberLLM

  • UberLLM is operated by AL Global LLC, a Wyoming limited liability company (ID 2025-001810253), 30 N Gould St, Ste R, Sheridan, WY 82801, USA. It is named on About, in the Terms (where "we" means that company) and in the site footer.

2026-10-30 — Where Claude, GPT and Gemini capacity comes from

  • How it works now says where Claude, GPT, Gemini and Grok requests run: through a seller's own account with a provider (in the past week: Amazon Bedrock, Anthropic, Google, OpenAI, OpenRouter, Snowflake, Venice AI, Bankr and Capminal), and what the market checks and does not check.
  • The seller rule is the same everywhere: list a provider key only if that provider's terms allow reselling it or giving others access; stolen, leaked, shared, free-trial and evaluation keys are prohibited.
  • Model pages say "Model by" and "Also sold directly by" instead of "Served by". The providers listed publish their own price for the model; they are not necessarily the sellers that serve your request.
  • The home page has a short "Where the capacity comes from" section. Its hero shows the median saving off list, not the deepest one. The Claude Code installer command moved to /claude-code, next to its SHA-256; the home page keeps the two environment variables.
  • Older changelog lines about a 5% credit fee now say it was the fee at the time.

2026-10-30 — Responses API: Codex CLI works with UberLLM

  • POST /v1/responses serves OpenAI's Responses API on every chat model, streaming and not: text, function tools, custom (freeform) tools, structured output, reasoning and usage with cost_usd. Same routing, retries, key limits and billing as chat completions. Responses API reference.
  • Codex CLI works: add a model provider with wire_api = "responses" in ~/.codex/config.toml (setup guide). Codex has called only the Responses API since version 0.95.
  • client.responses.create and client.responses.stream in the OpenAI SDK work.
  • UberLLM stores no responses: send the whole conversation in input each call. previous_response_id, conversation and background get a clear 400.
  • The Continue, Open WebUI, n8n, Goose and OpenCode pages no longer say the Responses API is missing. They still recommend chat completions, the setting each was tested with.

2026-10-30 — Fal video: priced at what it really costs us

  • A Fal video model can now cost less than the seller's discount says. Every 6 hours we compare what the marketplace charged us for it over 7 days with Fal's price for the same jobs. With at least 20 jobs, the model is priced at that share plus 5%, when that is lower. The listed price and the bill use the same figure. A figure more than 3 days old is not used.
  • If a video start gets no answer at all, we ask the same seller once more instead of a second seller, so one request can never start two paid videos.
  • The setup email's first video is now Kling 3 Standard, 5 seconds without sound, with today's price ($0.36).

2026-10-30 — Same reply shape from every provider, chat-only model list

  • A reply that calls a tool now always ends with finish_reason: "tool_calls" (stop_reason: "tool_use" on /v1/messages), streamed or not. Some Gemini models ended tool calls with "stop", so tools like Cline never ran them.
  • Replies keep only the OpenAI fields and our usage.cost_usd. Fields a provider added of its own (x_groq, service_tier, system_fingerprint, usage.queue_time and others) are gone.
  • Reasoning text comes in both reasoning and reasoning_content, with the same text, whichever provider served the call. reasoning_details passes through.
  • Requests that one provider refused now go through: thinking and other tuning fields it does not know are dropped for it, reasoning: {"effort": …} is sent as reasoning_effort, and content sent as a list of text parts goes as a string. A picture sent to a model that reads only text gets a plain 400 saying so.
  • GET /v1/models lists chat models only, so SillyTavern, LibreChat and n8n show models they can chat with. ?type=all lists every model; ?type=image, video, music, speech and transcription work as before.
  • Each chat model in GET /v1/models now has supports_tools, supports_reasoning, supports_vision, supported_parameters, and OpenRouter's architecture, top_provider and per-token pricing.prompt / pricing.completion.
  • Models that read pictures are marked so: Claude Sonnet 4.6, Claude Haiku 4.5, GPT-5 mini, Gemini 3.5 Flash, Kimi K3 and K2.6, MiniMax M3 and others were listed as text-only. Context and output limits now refresh every few minutes; Claude Haiku 5.5 shows its 1M context.
  • A model with many healthy providers no longer shows as unavailable after one request that no provider could take.

2026-10-30 — Video price per second, provider by provider

  • The Veo 3.1, Kling, Hailuo 02, Wan 2.7, Seedance and Sora 2 price pages (for example Veo 3.1) now list every resolution and sound option, with UberLLM's price next to fal.ai, kie.ai, Replicate and Google Vertex AI and the day each price was read.
  • "UberLLM lowest" appears only where UberLLM's price, with the 5.5% card fee added, is under every other price for the same clip. Where another provider is cheaper, the page says so. kie.ai's per-video prices are divided by the clip length to compare per second.
  • Prices refresh with the hourly rebuild. If a provider's list can't be read, the last price read is shown with the day it was read.
  • Video API pricing leads with the count across all providers, and its kie.ai table compares only prices a video is billed at exactly (the Fal models).

2026-10-30 — "Use with a cheaper API" guides with live model picks

  • Step-by-step pages with the exact settings for SillyTavern, Janitor AI, OpenCode, Kilo Code, Cline, Continue, Aider, Open WebUI, LibreChat, Chatbox (with a one-click import link), n8n (chat models and video jobs), the Vercel AI SDK and the OpenAI Python SDK and LangChain.
  • Each page suggests models for that tool with today's price, and its config names a model that is being served, re-checked every hour.
  • The Continue and Aider docs now match those apps' current versions: Continue needs useResponsesApi: false and useLegacyCompletionsEndpoint: false, and Aider needs a model settings file for diff edits.
  • Not covered: OpenAI's Codex CLI. Since version 0.95.0 it only calls the Responses API (/v1/responses), which UberLLM does not serve.

2026-10-30 — Your first API key is ready when you sign up

  • Every account now gets its first API key on its first visit to the dashboard, shown at the top with a request you can paste in a terminal. It stays there (also after you come back from the payment page) until you press I've copied it.
  • Before, you had to find Create key yourself, and some people who paid never made a key. Accounts that already have a key, or deleted all of theirs, get nothing new.

2026-10-30 — Apps ranking and "Connect UberLLM" for apps

  • New apps ranking: apps built on UberLLM, ranked by the tokens they send over the last 24 hours, 7 days or 30 days, with the models each one uses most. Model pages show the apps that use them.
  • To be listed, send HTTP-Referer (your app's address) and X-Title (its name) on your requests. OpenRouter's names work too (X-OpenRouter-Title, X-OpenRouter-Categories, X-OpenRouter-App-Visibility), so an app moving from OpenRouter changes only the base URL. Only the address is kept, never the page path or query.
  • "Connect UberLLM": send users to https://martllm.com/auth, they allow your app (with an optional spending limit), and your app gets a key on their own account with POST /api/v1/auth/keys. Same flow as OpenRouter's OAuth PKCE; only S256 is accepted. They can turn the key off any time on their keys page. See Apps.

2026-10-30 — Fal video prices follow the live Fal sellers

  • A Fal video model (ids starting fal-ai/) now costs Fal's own price less the discount of the cheapest Fal seller selling right now. Today that is 15% off. When a 25%-off seller is selling again, prices drop back by themselves.
  • Since 10-29 the site listed 25% off, but no seller was selling at 25% off, so every job ran at 15% off. We covered the difference on every job already made. Nobody is charged again for them.
  • If the cheapest seller turns a job down, a dearer one makes it. You pay that seller's price, shown as estimated_cost_usd on the reply, and never more. A failed job still costs nothing. Stopping a running job is best effort: if the clip was already made, it is billed.

2026-10-30 — n8n video templates, n8n node and ComfyUI nodes

  • Five n8n workflows that make videos with one key, to import by URL: Veo 3.1 Fast text to video, Kling 3.0 image to video, a faceless Shorts pipeline, a UGC-style product ad, and a drop-in for the Gemini "Generate a video" node that stopped working when Google turned off the Veo 3.1 preview models in the Gemini API on 22 October. Each was run end to end in n8n 2.45. What a run costs today: n8n video templates.
  • Source for an n8n community node (n8n-nodes-uberllm: Chat, Generate Image, Generate Video, Get Job) and ComfyUI nodes (comfyui-uberllm: text and image to video, image, image edit).

2026-10-30 — Setup guides for 14 apps

2026-10-30 — "What is the cheapest API?" for every model

  • New page for each chat model, for example the cheapest DeepSeek V4.1 Flash API: UberLLM's price next to the same model's price at OpenRouter, Vercel AI Gateway, Requesty, DeepInfra and Novita AI, read from each one's public model list every hour, with links to those lists.
  • Video models too, for example the cheapest Kling 3 Pro text-to-video API: each resolution and sound setting next to fal.ai's own price and kie.ai's price for the same clip.
  • A page exists only when UberLLM is lower than every other provider listed, on input and on output (for video, on every clip), even with the 5.5% card fee added. Otherwise there is no page. All of them are under Answers, and each model page links to its own.

2026-10-30 — A model nobody serves still answers

  • When nobody is serving the model you asked for, the closest working model answers instead of an error. It is chosen from models that are up right now: uncensored only when you asked for an uncensored model, with tool support when your request has tools, big enough for your prompt, at the same price or lower when there is one (otherwise the cheapest). The reply's model field names the model that answered, and you pay that model's price.
  • Want only the exact model? Send your own "models": [...] list, or set provider.max_price to cap what a call can cost.

2026-10-30 — Unavailable models stay marked unavailable

  • A model with no provider serving it now shows "available": false until a provider serves it again. Before, it could switch back to true after a few quiet hours while every call still failed. When a provider returns, the model shows as available again within a minute, by itself.
  • The "These models work now" suggestions in an error only name models that are answering.
  • Never wait on one model: add "models": ["model-a", "model-b"] to a request and, if the first has no provider, the next one answers. The reply's model field says which one did. See Models and routing.
  • Dashboard → Activity and its CSV export: a request your app stopped after the provider had already started work now shows what it cost. Before, it said "You were not charged" even when it had a cost.

2026-10-29 — Every model can be used: billed what the provider charged

  • Some models were refused every time with "No one is selling this model at its listed price" because every provider asked more than that price. They now answer, at the provider's price. DeepSeek V4.1 Flash Uncensored is one.
  • A model that nobody is selling right now shows "available": false everywhere (GET /v1/models, GET /v1/models/<id> and the website) and stays that way until a provider serves it again. Before, the list and the single-model answer could disagree. Today that is GLM 5.3 Uncensored, GLM 5.3 Flash Uncensored and MiMo V2.6 Flash Uncensored: no provider is offering them at the moment.
  • Any provider that is up can now serve a call, for chat, images, video, music and speech.
  • You pay what the provider charged for each call. UberLLM still adds no fee. Each call is listed with its cost in Dashboard → Activity.
  • The listed price now follows the market. When nobody sells a model at its usual price, it shows the cheapest provider's real price, and it goes back down by itself when a cheaper provider returns. It is what a call typically costs, not a limit.
  • available in GET /v1/models now means a provider can serve the model right now.
  • Want a hard limit? provider.max_price (chat) and max_cost (images, video, music, speech) still work: a call is never billed above them. A call that fails is still not billed. See Pricing.

2026-10-29 — Uncensored text models are listed

  • Uncensored chat models such as DeepSeek V4.1 Flash Uncensored, GLM 5.3 Uncensored and Venice Uncensored now show in the model list and the playground without turning anything on.
  • Adult picture, video and sound models stay hidden until you tick Show adult models.
  • The API is unchanged: GET /v1/models still marks these models "nsfw": true, so you can leave them out of your app.

2026-10-29 — Your credit is the exact amount you pick

  • Adding credit now works like this: you pick the credit, and a card fee of 5.5% + $0.50 is added on top when you pay. $10 of credit costs $11.05, $50 costs $53.25. Until this change, a 5% fee came out of the amount, so paying $10 added $9.52.
  • Dashboard → Credits shows the credit, the card fee and the total before you pay. Requests still carry no fee.
  • Credit you already bought is unchanged. A payment page opened before this change still adds what it showed you.

2026-10-29 — Every Fal video price next to fal.ai and kie.ai

  • New page martllm.com/fal-api-pricing: every Fal video model on UberLLM (Kling 3.0, Kling O3, Kling 2.6, Veo 3.1 Fast, Wan 2.7, Hailuo 02, Vidu Q3, LTX 2.3) with each resolution and sound option, priced per second next to fal.ai's own price and kie.ai's price for the same clip. Prices are read live. Veo 3.1 Fast and Hailuo 02 show fal.ai and UberLLM prices only.
  • The home page now opens with two choices: Start saving (call discounted models) and Start earning (list your API credits).

2026-10-29 — Short answers with live prices on every model page

  • Every model page now opens with a short answer to "How much does it cost?": the price now, how far it is below the list price and how many sellers can serve it. Each number comes from the live market; a number we don't have is left out.
  • New answer pages at martllm.com/answers: the cheapest model in each family (Claude, GPT, Gemini, DeepSeek, Qwen and more), video price per second (Veo 3, Kling, Wan, Sora 2 and more) and image price per image. They are rebuilt every hour.
  • Sell now answers "Where can I sell unused AI capacity?" with how many sellers, offers and models are on UberLLM today.
  • A link to martllm.com pasted in Discord, Slack, X or iMessage now shows a dark card with live prices for a video, an image and a chat model, and a "Start in a minute" button.
  • Every docs page has its own card and text. A link to this changelog shows the three newest changes instead of the homepage card.

2026-10-29 — Model names: a short name works, and a wrong one says what you meant

  • A model name without its vendor now works when exactly one model has that name: grok-4.6 is x-ai/grok-4.6. If two vendors use the same name, we still ask you to say which one, and never pick for you.
  • uberllm/ in front of a model id (as some tools add) now works for images, video, music and speech too, not only for chat.
  • An unknown model name now ends with "Did you mean: …?" and up to three close model ids (for video, only video models). This works the same on /v1/chat/completions, /v1/messages and every media endpoint.
  • When a request is turned down with only "Invalid request parameters", the error now says what that means: nothing was charged, which fields models most often refuse (tool definitions, images or files, response_format, sampling settings), and that support can look it up from the reply's X-Request-Id.

2026-10-29 — DeepSeek V4.1 Flash Uncensored works again, at the price its sellers ask

  • Every seller of deepseek/deepseek-v4.1-flash-uncensored asks more than its reference price ($0.045 in / $0.50 out per 1M tokens), so every call to it was turned down. It is now listed at the cheapest seller's price (today $0.15 in / $0.60 out per 1M tokens) and calls are served. You are billed what the seller charges, never more than the listed price.
  • When a seller offers it at or under the reference price again, the listed price goes back down by itself.

2026-10-29 — Video pricing page: what you get first, prices second

  • /video-api-pricing now starts with what making video on UberLLM gives you: one key for chat, images and video; a video that fails is not billed; and video straight from Claude Desktop or Claude Code through the MCP server (start_video, then wait_job). It names Kling 3.0 Standard as the model to start with. The model-by-model price table against kie.ai is unchanged.
  • /docs/video and /docs/videos now open the images, video, music and speech guide instead of a missing page.

2026-10-29 — A model you can't buy right now says so, and names ones you can

  • A model is now shown as available only when at least one seller offers it at or under our listed price. Before, a model whose every seller had raised their price above ours (DeepSeek V4.1 Flash Uncensored today) still showed as available, and every call to it was turned down. It now shows as unavailable, best-price: and fastest: skip it, and it comes back by itself as soon as a seller is at our price again.
  • "No one is selling this model at its listed price" and "No one could serve this model" now end with up to three chat models that work right now, for example "These chat models work now: google/gemma-4-uncensored, …". We never switch the model for you; to have backups tried automatically, list them in models.
  • A call that is not streamed and fails after a few seconds now gets its real error status (for example 503). Before, after 5 seconds it could come back as HTTP 200 with the error in the body, which most SDKs read as an empty answer. Very slow calls still get a keep-alive after 45 seconds, so clients that give up at 60 seconds stay connected.

2026-10-29 — MCP: no more 6-minute "calls", every tool call in Activity, model list fixed

  • Connecting from Claude Desktop or Claude Code used to leave a request open for about 6 minutes (sometimes an hour) every time. It looked like a tool hanging, but it was only the app waiting to hear whether our tool list had changed. The list never changes, so we no longer ask apps to wait for that. Tool calls themselves were never slow.
  • Every image, video and quote request a tool makes is now tagged via: mcp in your Activity. A request turned down before it reached the model is listed there too, with the reason, so a failed tool call always leaves a trace.
  • GET /v1/models listed 21 video models (Kling on Fal) as also taking /v1/chat/completions and /v1/messages. A chat request to them was always turned down. They now list only /v1/video/generations, and best-price:chat / fastest:chat can never pick a model that makes only pictures, video, music or speech.

2026-10-29 — Video: we name the models that work when yours can't run

  • When nobody can make a video with the model you picked, the answer now ends with the video models that work right now, for example "These video models work now: fal-ai/kling-video/v3/standard/text-to-video, …". Models with videos finished this week come first, then the ones closest in price. The same goes for images, music and speech. We never switch the model for you (another model is another price and another look); to have it tried automatically, list backups in models.
  • The same answer comes back when a video job was accepted and then lost by every seller, after our automatic restarts. You are not charged for it.
  • A model that just lost a video now shows as unavailable in GET /v1/models for up to 10 minutes, and best-price: / fastest: skip it. It comes back as soon as a video on it succeeds.
  • Sending audio to a video model that has no sound setting (for example google/veo3-fast-text-to-video) no longer fails the video. The field is left out and the model makes the clip as it always does.

2026-10-29 — Send video to Gemini, and fewer refused calls

  • Gemini models now watch video, read PDFs and listen to any audio format you send in a chat. Send a video as video_url, a document or clip as file, or audio as input_audio in m4a, mp4, ogg, flac or aac. Before, Google turned these down with "Invalid content part type" or "Invalid audio format". Plain links and data: links both work.
  • Tool loops on DeepSeek work with the plain OpenAI SDK. Some DeepSeek providers refuse a tool call unless every earlier reply's thinking is sent back, which the SDK never does. We now answer that call with thinking off instead of failing.
  • reasoning_effort no longer gets a call turned down. When a model doesn't offer the level you asked for (for example "none" or "minimal" on GPT-6.1 Sol), you get the nearest level it does offer. When it can't combine thinking with function tools, the call runs with tools and the model's own thinking setting.
  • When max_tokens plus your prompt is more than the model can hold, the reply gets all the room that is left instead of an error.
  • Image edits: size now works on every edit model. Models that set their own resolution (Qwen Edit Uncensored, Seedream 5 Lite Edit, FLUX.2 Max Edit) get the nearest aspect ratio instead. Before, they refused the edit with "Resolution is not supported".
  • When an image or video costs more than your max_cost, the message now shows two different numbers (for example "up to $0.0401, above your max_cost of $0.0400"). Before, both could read "$0.04".

2026-10-29 — We find out when a call of yours fails, without you telling us

  • Every 5 minutes we now check the calls and jobs of every account that has added credits. When one fails for a reason that is ours to fix (a provider turned down something we should have sent another way, nobody was selling the model, a setting the model does not take), our team is told within minutes, once per kind of problem, with the model and the error you saw.
  • We are also told when a model fails most of its recent calls or jobs, even when our model list still shows it as available. That way a model that is down gets fixed or taken off the list before many people try it.
  • If your key is connected but no call or job has come from it yet, or you added credits a day ago and nothing has worked, we are told too, so a person can reach out and help you get started.
  • Failures you can fix yourself, where our error already says what to change (low balance, a request over 10 MB, a model id that does not exist), are not reported.

2026-10-29 — Long prompts start answering sooner

  • Prompts over 128k tokens: when the first provider is slow to start, we also ask a second provider, and the first to answer wins. That second request used to go back to the same provider. It now goes to a different fast one. Live: a 203k-token prompt now starts answering in about 4 seconds; before, prompts this size waited 9–18 seconds on DeepSeek V4.1 Flash.
  • Prompts of 8k–32k tokens: a provider that turned down a call once could be left out for a whole day, even though it was the fastest one for that size. Now it is tried again within 15 minutes. One fast answer is enough for it to stay in the running.
  • We now check speed separately for short, ~20k, ~60k and ~150k-token prompts every 5 minutes, so a slowdown at one size is caught even when short calls are fast.
  • On /v1/messages, asking for thinking now returns the model's thinking on every provider. Some providers (for example the ones serving DeepSeek V4 Flash) speak only the OpenAI format; their thinking was dropped when we turned the reply into Anthropic's format, so you paid for it and saw nothing. Without thinking in the request, replies are unchanged.

2026-10-29 — Replies finish sooner, and no hidden thinking you did not ask for

  • Every reply now ends as soon as the model is done. Before, a reply could wait 1–2 seconds (sometimes up to 15) at the very end while we looked up what the call cost. On DeepSeek V4.1 Flash that wait was on almost every call, so an agent that makes dozens of calls in a row feels it the most. The lookup now happens after your reply. Your balance and Activity still show the exact charge. When the provider has not priced the call yet, usage.cost_usd in the reply is our estimate.
  • On /v1/messages, models think only when you ask for it, as on Anthropic's own API. Send "thinking": {"type": "enabled", ...} or "adaptive" to get it. Before, DeepSeek and other reasoning models always thought first, and with a small max_tokens the thinking could use the whole budget and return an empty answer. Claude models already worked this way and are unchanged.
  • A request one provider's own content filter turns down ("may contain inappropriate content") now goes to another provider for the same model instead of failing. One provider was turning down about 230 DeepSeek V4.1 Flash calls a day that others serve fine.
  • A provider whose account was suspended no longer fails your call. The next attempt goes to a working one. This was about 730 failed calls a day on GLM 5.2 and DeepSeek V4.1 Flash 1017.
  • A route that stalls or fails now loses its place in the fast-route ranking. Before, a route that had been fast kept getting picked even while it stalled, and some calls waited up to 10 seconds before moving on.
  • Streamed chat completions always end with data: [DONE]. Gemma 3 27B ended without it, so clients that wait for it hung after the last word.

2026-10-28 — Claude Code no longer stops on some DeepSeek replies

  • Some providers sent DeepSeek's reasoning without the field Claude Code and the Claude Agent SDK require, so a long session could stop with "Missing required field in assistant message: 'signature'". Replies now always include it. Nothing to change on your side.

2026-10-28 — Long prompts start answering up to 10x sooner

  • Every request now goes to the cheapest provider that is fast enough for its size. We measure each provider's first-reply time separately for short and long prompts, because the cheapest one is often the slowest on long ones.
  • DeepSeek V4.1 Flash with 32k–128k-token prompts: first word in 3.3s typical (was 18s). Above 128k: 3.4s (was 40s).
  • You never pay more than the listed price. When a faster provider costs more, we cover the difference.

2026-10-28 — See what you saved, and share it

  • Your dashboard now shows what your calls would have cost at each model's official list price, and what you paid instead. Video and image calls count at what you paid, so the number never overstates.
  • Share makes a link (martllm.com/s/…). Pasted into X, LinkedIn, Slack or Discord it shows your card as an image. The card holds only the numbers and model names: no name, email or prompts. People who sign up from it count under Invited on your Share page.

2026-10-28 — Videos that hit a busy moment are retried for you

  • When no provider is free to make a video or music track, or one is too busy (several jobs sent at once), UberLLM now starts the job again by itself, up to two more times within 10 minutes. You keep the same job id and the same most-it-can-cost; the job only takes a little longer. Before this, those jobs ended failed and you had to send them again.

2026-10-25 — One account per person; Terms and Privacy

  • A network can open at most 3 new accounts a day. Signing in to an account you already have is never limited.
  • An inbox gets one account: name+tag@… and, for Gmail, dotted variants of an address that already has an account are told to sign in with it.
  • Terms of Service and Privacy Policy are published. In short: we don't store your prompts or answers, only what we need to bill you.

2026-10-25 — Fewer refused requests

  • A chat request is no longer refused when the fastest seller is above the listed price but others are not: it goes to the rest of the market at the same price.
  • Model ids with a provider prefix in front (uberllm/<model>, as LiteLLM and OpenRouter-style configs send them) now work.
  • A request over 10 MB answers 413 request_too_large instead of "not valid JSON".
  • Pages no longer say prices include a fee: requests carry no UberLLM fee.
  • A video or image whose size the seller never measured no longer shows "bytes": 0; the field is left out.
  • Models that trade at more than 98% off (Qwen3 VL 235B among them) are no longer refused with "no one is selling at its listed price" while sellers are serving them. You are still billed what the seller charged.
  • A new API key now comes with a ready-to-paste curl request, so you can check it works in one step.
  • /pricing, /changelog, /keys and /register now open the right page.

2026-10-25 — 21 Fal video models, 15–25% under Fal's price

  • Wan 2.7, Kling 3 (Standard and Pro), Kling O3, Kling 2.6 Pro, Veo 3.1 Fast, LTX 2.3, Hailuo 02 and Vidu Q3 — text-to-video and image-to-video — are now on /v1/video/generations under their Fal ids (fal-ai/…).
  • They cost 15–25% less than Fal charges for the same clip, and UberLLM adds nothing. A 15-second 720p Wan 2.7 video costs $1.125–$1.275; the same model through the marketplace's other sellers was $6.60.
  • The price is fixed by what you ask for, so estimated_cost_usd on the reply is the exact bill. Failed jobs cost nothing.
  • See Fal video models.

2026-10-22 — Holds cover tools and images on chat completions

  • The amount held while a /v1/chat/completions request runs now counts the whole prompt: tool definitions (tools, functions), tool calls and their arguments, tool results, response_format schemas, and image, audio and file parts. Before, it counted message text only. /v1/messages uses the same measure. You are still billed only what the provider counted.
  • An image is held by its size when it is sent inline (PNG, JPEG, GIF, WebP). An image sent as a link is held at the most the model bills for one picture: about 5,200 tokens on Claude models and 35,700 on the others, released when the request is billed.
  • /v1/messages/count_tokens now includes images and documents.

2026-10-22 — A call the provider charged nothing for costs nothing

  • Providers bill in whole millionths of a dollar, so a very short call can cost them $0. That call now costs you $0 too. Before, it was billed at the model's listed price for its tokens (under $0.000002 a call).
  • Docs: the quickstart, routing and SDK pages no longer mention the old 10% fee; adding credit is by card.

2026-10-22 — No UberLLM fee; prices read like the marketplace

  • UberLLM no longer adds 10% to requests. Text, images, video, music and speech cost exactly what the provider charged, and never more than the model's listed price. Adding credit kept its 5% fee at the time. (Since 2026-10-29 the card fee is 5.5% + $0.50, added on top; there is no fee on requests.)
  • "Price now" and "% off" on the site and model pages are the marketplace's own numbers: the lowest price a provider sells the model at right now, and how far that is below the list price. Video rows read "$X / job" as the marketplace writes them; that is the rate per second of video at the lowest resolution, so a 5-second job costs 5× it.
  • A request is now only sent to a provider whose price is at or below the model's listed price: chat, /v1/messages, images, video, music, speech and transcription. When nobody sells the model at that price right now, it answers 503 no_available_provider and nothing is charged. available in GET /v1/models and the best-price: / fastest: aliases follow the same rule.
  • Chat replies (streamed or not, OpenAI and Anthropic formats) no longer carry the provider's own fields, such as provider, cost and cost_details. What a call cost you is usage.cost_usd.
  • A source picture link that redirects (such as https://picsum.photos/…) is followed to the picture before the job starts. A link that does not lead to a picture is refused with 400 invalid_image_url, before anything is held.
  • A video or music job's max_cost_usd is the amount held for it, on every state. Before, a finished job showed a different, lower figure.

2026-10-22 — Cache-write prices; holds that cover small requests

  • GET /v1/models lists pricing.cache_write_usd for models whose providers charge for writing the prompt cache (Claude: 1.25× the input price). Prompt tokens written to the cache are billed up to that price, the same way the provider bills them. Before, they were counted at the input price.
  • The amount held while a chat request runs now allows for the provider's own chat template (a short message can be billed 30 or more prompt tokens) and for reasoning models that think past a small max_tokens. A hold is released when the request is billed; you only pay the bill.

2026-10-21 — Quotes show applied defaults; Claude bills on the real counts

  • A quote ("validate": true or POST /v1/quote) lists defaults_applied: the settings you left out that the seller needs, with the value sent (a reference-to-video model gets aspect_ratio 16:9).
  • A Claude stream whose seller reports no token counts is now priced on the upstream's own counts, not our estimate, before the listed-price ceiling is applied. The money held while a Claude request runs is sized for Claude's own tokenizer, which counts more tokens than ours.

2026-10-21 — Speech aliases fit the voice; request ids on media errors

  • best-price:speech and fastest:speech fit your voice to the model they pick. Voices are each model's own, so OpenAI's default alloy was refused by every speech model. Now a voice the model has is kept, alloy maps to a model's af_alloy, and any other voice gets the model's first voice. Naming a model keeps your voice as sent.
  • A failed image, video, music, speech or transcription request now answers with X-Request-Id, the id of its Activity row. Before, only some refusals had it.

2026-10-21 — Costs on chat replies, aliases that always work, media fixes

  • Chat replies say what they cost. /v1/chat/completions and /v1/messages answers carry usage.cost_usd: what the request is billed, in US dollars, our fee included and never above the listed price. In a stream it is in the last usage chunk (message_delta on /v1/messages). An answer that is not streamed also has the X-UberLLM-Cost-USD header, unless it was slow enough to start with keepalive spaces; then the body field is the only place.
  • When a model's seller streams no token counts (DeepSeek on /v1/messages), the final message_delta shows the counts the request was billed for, with output counted from the text, thinking and tool input we sent you. Before this fix, output on those streams was counted as 0.
  • When a stream carries no token counts, the request is billed and listed in Activity with the provider's own token counts (prompt, cached and output) whenever we read its charge afterwards, or whenever a price limit would otherwise be checked against our estimate. Before, a long agent prompt could show far fewer prompt tokens than it really had.
  • A /v1/messages request's hold now counts tool definitions, tool calls, tool results and earlier thinking, as /v1/messages/count_tokens already did. An agent's prompt is mostly tool results; a hold sized from the text alone could be smaller than the request's real cost.
  • best-price: and fastest: pick only models that are available right now: the live market shows a healthy seller, and our last request to the model in the past 10 minutes found one. On a media route, when the model an alias picked turns out to have nobody taking requests, the request moves on to the next model the alias stands for (up to 3). When no model for the job is available, the answer is 503 no_available_provider.
  • A model sent to the wrong route answers 400 wrong_endpoint naming the right one, before anything runs or is charged: an image, video, music, speech or transcription model on a chat route, a chat model on a media route, or speech and transcription swapped. Activity lists it as refused, with no tokens.
  • Every model on /v1/models, /v1/models/{id}, the public models list and the live board says available (can it be used right now), nsfw (made for adult or unfiltered content) and the right input_modalities: models that take a picture list "image", the video upscaler lists "video".
  • A video job's estimated_cost_usd is its listed price, the same as validate and /v1/quote cost_usd. max_cost_usd is still the most it can cost.
  • The final cost of a finished job is read within seconds (it was checked every 15 s, and the report's 47 minutes was when it was looked at again: the charge had landed 18 s after the video). callback_url fires once, when the cost is final. After 30 minutes without a charge, the job is billed at its listed price.
  • Video models priced per job that list clip lengths are priced per second, as they are billed (Kling: its listed price is per second, so a 5 second clip costs five times it).
  • Reference-to-video models default aspect_ratio to 16:9 (shown as the default in the model's schema). A video model that needs a picture answers 400 image_required without one, before anything is held.
  • Transcription reads the audio format from the file itself, so an mp3 named recording.out works.
  • response_format defaults to url on /v1/images/generations and /v1/images/edits. A picture the model sends inline is kept 24 hours as a file of your account and returned as a link; "response_format": "b64_json" returns it inline.
  • Refused image, video, music and speech requests now show in Activity.
  • /v1/quote answers with both id and quote_id (the same value).

2026-10-21 — Model availability on the website, Models in the dashboard

  • A model's page says "Not available right now" when no provider serves it. Requests to it are refused until one does, so the page shows no price or code example for it. Such models are left out of the model list, the Playground and related-model links.
  • The dashboard has its own Models page with the same list and prices. Image, video and music models have a "Try in Playground" link that opens the Playground with that model picked.
  • Adult models are hidden from the model list and the Playground. "Show adult models" shows them; the choice is kept in your browser. Their pages still open from a direct link.
  • Video prices billed per second also show the price of a 5 second clip.

2026-10-21 — Never billed above the listed price

  • A request is never billed above the listed price for the model. Prices include UberLLM's 10% platform fee (ended 2026-10-22). A chat call cost what the provider charged plus 10%, but never more than its tokens at the listed price (cached prompt tokens at the listed cache-read price); when the provider charged more, UberLLM pays the rest.
  • The same holds for images, video, music and speech: a job is never billed above the listed price of what you asked for, which is the cost_usd of its quote. max_cost_usd is only what is held while it runs.
  • A seller's own offer is still billed at that offer's listed price.

2026-10-21 — Quotes at the shown price, playground and MCP fixed

  • A quote (POST /v1/quote or "validate": true) now says what the request is expected to cost at the price shown on the model's page (cost_usd), next to the most it can cost (max_cost_usd, what is held while it runs). Both answers are "object": "quote".
  • The dashboard playground and every MCP tool call work again (they were refused since the 10-21 release).
  • GET /v1/models?type=… lists a media model only when it can be bought now (live and priced), the same models the board and its counts show.
  • A key's daily limit can't be more than its spending limit; the Keys form shows it only under a monthly or all-time limit.

2026-10-21 — Callbacks and alert webhooks

  • callback_url on a video or music job: when the job ends (succeeded and billed, failed, cancelled or expired) we POST the job's JSON to your https link. The job shows how it went in callback. See Callbacks.
  • Every callback is signed: UberLLM-Signature: t=…,v1=… (HMAC-SHA256 with your account's webhook secret), with a stable UberLLM-Event-Id. Failed sends are tried again after 30 seconds, 2, 10 and 30 minutes and 2 hours. The secret is under Dashboard → API keys, and can be replaced there.
  • The low-balance and held-money alerts can also go to a webhook: Dashboard → Credits → For developers, or alerts_webhook_url on PUT /api/v1/billing/alerts.

2026-10-21 — Held quotes, fallback models, uploads, tags, daily key limits

Speech

  • Text to speech and transcription take the same validate, max_cost, metadata and provider.max_price options as images (form fields on a transcription), speech also takes models: [...], and best-price:speech / best-price:transcription pick the cheapest live model. Every answer carries X-UberLLM-Cost-USD and X-Request-Id headers; a transcription's JSON also has usage.cost_usd.
  • GET /v1/models?type=speech, ?type=transcription or ?type=audio lists them.

Prices and routing

  • POST /v1/quote prices an image, video or music request and holds that price for 15 minutes: send the same request with quote_id and it is billed at most max_cost_usd. One quote starts one request. See Make a video or image.
  • models: [...] names fallback models, tried in order only when the one before cannot be served right now. Works on chat, image, video and music.
  • best-price:<job> and fastest:<job> model names pick the cheapest, or quickest measured, live model for chat, image, image-edit, text-to-video, image-to-video or music. GET /v1/models/aliases lists what each stands for now. fastest: needs at least 5 timed requests per model in the last 7 days.
  • Chat provider options now act: allow, deny and order name the providers the marketplace may use, sort picks price, latency or uptime, and max_price is a hard ceiling on what you are billed per million tokens. An option that cannot be applied is refused with 400 unsupported_routing_option instead of being ignored. See Models & routing.
  • Image, video and music take provider.max_price per unit.

Images and videos

  • POST /v1/files uploads a picture once (10 MB, kept 24 hours); use its id or url anywhere a source picture goes.
  • output_upload_url on a video or music job uploads the finished file to your own storage (a presigned https PUT link). The job shows how the upload went.

Tags

  • metadata (up to 16 text values) on any request is stored with it, shown in Activity, filterable with ?tag=key:value, and included in the CSV export.

Keys and alerts

  • API keys can have a daily limit (daily_limit_usd) on top of their spending limit. A request must fit under both, and the daily limit resets at midnight UTC. A request over either limit is refused with 402 key_spend_cap_exceeded and nothing is charged. Set it on the Keys page or with PATCH /api/v1/keys/{id}.
  • On the Credits page you can ask to be emailed when your balance falls below an amount you choose. It is off by default, sends one email each time the balance drops below that amount, and goes to the account's owners.
  • If money set aside for an image, video or music job is still held 15 minutes after the job ended, the account owners get an email listing those jobs and our team is told. That money is still yours and will be released.

2026-10-21 — Prices you are billed, MCP server, media playground

Prices

  • Every price on the site and in /v1/models is now the rate we bill, our 10% fee included at the time (that fee ended on 2026-10-22), from the same numbers billing uses. Before, the board showed the cheapest seller asking price, which could be far below what a request was billed.
  • "% off" compares that billed rate with the model maker's own list price. A model under 1% off shows no discount.
  • A request is still billed what the provider charged plus 10%, so one call can come in a little above or below the shown rate. See Pricing.

MCP server

  • https://api.martllm.com/mcp lets Claude Code and other MCP clients list models, price a request, make images, start videos and music, wait for a job and check the balance, with your API key. Setup: claude mcp add --transport http uberllm https://api.martllm.com/mcp --header "Authorization: Bearer $KEY". See MCP server.

Dashboard

  • Playground: send one prompt (and a start picture if you like) to 2 to 4 image, video or music models, see the most each can cost, then run them side by side and compare the result, how long each took and what each cost. Runs are charged to the key you pick and show in Activity.

2026-10-21 — Image edit inputs, media costs, per-project Claude Code setup

Images and videos

  • /v1/images/edits now takes the source picture as image (an https:// link, a data: URI or plain base64), image_url, input_images, images, or a multipart image upload. Video takes a start frame in image_url and reference pictures in input_images. Make a video or image lists the field for every route, with one working request each.
  • Every image answer and every finished job carries what it cost in usage.cost_usd and in the X-UberLLM-Cost-USD header.
  • "validate": true checks a media request and says the most it can cost, without charging. "max_cost" refuses anything that can cost more than you set.
  • Idempotency-Key on a video or music job: sending the same request again within 24 hours returns the same job instead of a second charge.
  • A clip length, resolution or shape the model does not make is refused up front, with the ones it does make.
  • Job status is always one of queued, running, succeeded, failed, canceled or expired. Poll every 5–10 seconds (callback_url came later, on 2026-10-21).
  • How long results are kept is now written down: about 3 hours for video and music, about 15 minutes for image links.

API

  • GET /v1/models lists image, video and music models too, each with type, endpoints and pricing; filter with ?type=chat|image|video|music.
  • GET /v1/models/<id> gives one model, with a JSON Schema of the request body for media models.
  • GET /v1/balance returns balance_usd, held_usd and available_usd.
  • Every /v1/* answer carries X-RateLimit-Limit, X-RateLimit-Remaining and X-RateLimit-Reset; a 429 from the key's limit adds Retry-After.
  • When a seller's own provider account fails, UberLLM tries other sellers and, if none is left, answers 503 no_available_provider. A seller's error no longer reads as if your key were wrong.

Claude Code

  • The installer can set up one project only: sh -s -- --project writes .claude/settings.local.json and leaves your other projects alone.
  • The key is no longer written into a settings file: it is saved to a file only you can read and Claude Code reads it through apiKeyHelper (macOS and Linux).
  • Profiles now also set CLAUDE_CODE_SUBAGENT_MODEL, so subagents run on the model you chose.
  • New: your UberLLM balance in Claude Code's status line (curl -fsSL https://martllm.com/install/statusline.sh | sh -s -- --install).
  • The Claude Code guide now covers which setting wins, mapping every model slot, context and compaction, what changes when you leave the claude.ai login, and where your code goes.

Models table

  • Each model shows how many sellers can take a request right now, out of all the sellers listing it.
  • The count line names both numbers: every model in the catalog, and how many the live market quotes right now.
  • Each model's page shows its type, a copy button for its id, the sellers that are healthy now, and a working request for picture edits, image-to-video and reference-to-video models.

2026-10-21 — See what is running, why a request failed, and what each key spent

  • In progress: video and music jobs that are still running show on Overview and Activity with their model, time so far, progress and the amount held, and a Cancel button. The "held" line on your balance now links to that list.
  • Failed requests say why: each failed row in Activity shows the reason in plain words and whether trying again may help. Click any row for its details and its job id.
  • Filters and per-key totals: filter Activity by period, key, model and result; see what each key spent; download what you see as CSV.
  • API keys: delete a key, and edit its name, spending limit and request limit at any time. A change applies to the key's next request.
  • Models is now in the dashboard menu, and each Activity row links to its model.
  • Credits: payments made through Stripe have a Receipt link next to them.

2026-10-21 — Speech: voice-overs and transcripts

  • Text to speech: POST /v1/audio/speech turns text into an audio file, billed per started 1K characters.
  • Speech to text: POST /v1/audio/transcriptions turns a recording into text (or subtitles), billed per minute of audio.
  • Same key and balance as everything else, and both work with the OpenAI SDK. A failed request costs nothing. See Images, video, music and speech.

2026-10-21 — Every price next to the official one

  • The models table shows each model's official price beside ours, and % off is now what you are billed against that official price. It used to show the market's own figure, which could read 100% off.
  • Videos priced by resolution show from $X per second (the 480p rate), ours and the official one, instead of calling a per-second rate a price per video. The kie.ai comparison prices those models for the same 5-second 480p clip.

2026-10-21 — A clearer home page

  • The home page describes what UberLLM is: an API marketplace for AI models.
  • The home page shows sample clips generated on UberLLM with current video models, and the models table opens on video. A pricing section shows the largest current % off list price, that failed requests are not billed, and the credit purchase fee (5% at the time). (Since 2026-10-29 the card fee is 5.5% + $0.50, added on top; there is no fee on requests.)

2026-10-21 — Add credit without leaving the dashboard

  • Dashboard → Credits now accepts card payments on the page through Whop's checkout. Cards, Apple Pay and Google Pay are supported where the device supports them, and local currencies where Whop offers them. Credit is added within a minute.
  • A refunded or charged-back payment takes back the credit it added, up to what is left of it. The fee on credit purchases stayed at 5% at the time. (Since 2026-10-29 the card fee is 5.5% + $0.50, added on top; there is no fee on requests.)

2026-10-21 — A 10% UberLLM fee on requests (ended 2026-10-22)

  • Every request (text, images, videos and music) now includes a 10% UberLLM platform fee on top of the provider's price. Listed prices include the fee. % off list is unchanged. See Pricing.
  • The fee on credit purchases stayed at 5% at the time. Sellers still pay no fee on their sales. (The 10% request fee was removed on 2026-10-22. Since 2026-10-29 the card fee is 5.5% + $0.50, added on top.)

2026-10-20 — A live market on the home page

  • A single models table on the home page. Prices, % off, liquidity, 24h activity and offers refresh every 30 seconds, with market totals above the table: average savings, money saved in 24h and active offers. Every column is sortable.
  • Tabs for Text, Image, Audio, Video and Decisions. Old /models and /market links open the table.
  • A savings calculator estimates a monthly bill at current prices.

2026-10-20 — Images, videos and music

  • Make images (/v1/images/generations), videos (/v1/video/generations) and music (/v1/music/generations) with the same key and balance. See Make a video or image.
  • Billed at the provider's cost of what was made, with no markup at the time. Failed, canceled and expired jobs are not billed. Jobs show in Dashboard → Activity.
  • Browse them at martllm.com, each with its price per video, picture or track.

2026-10-19 — Single marketplace routing, billed at the seller's price

  • Every request is now sent to a marketplace that selects a seller for each call. Each request is billed at the price that seller charged, with no inference markup at the time. The fee on credit purchases stayed at 5% at the time. (Since 2026-10-29 the card fee is 5.5% + $0.50, added on top; there is no fee on requests.)
  • Automatic retries: a call that fails before any content arrives is retried (up to 4 attempts within 45 seconds). A model nobody is serving right now returns 503 with the model named.
  • :floor / :nitro and the provider object are still accepted, but no longer change price or speed. See Models & routing.
  • GET /v1/models lists only models that can be served right now.
  • New accounts start at $0; there is no sign-up credit.

2026-07-24 — Launch

UberLLM launched as a two-sided marketplace for LLM inference: one API key for the catalog, with requests routed to healthy providers and weighted toward lower prices.

  • Gateway: OpenAI-compatible /v1/chat/completions and Anthropic-compatible /v1/messages with SSE streaming, automatic failover, and health-aware routing (providers with recent outages are deprioritized, and retries are bounded).
  • Market: 380 models on the book — live-priced curated routes plus external market asks to beat. Modality tabs, capability filters, per-model 24h activity, liquidity, and offer counts at /market.
  • Selling: list any OpenAI-compatible endpoint at your own price (absolute or a multiplier of the reference), pass a one-shot validation probe, earn on every routed request. Health is scored passively from real traffic, without probe requests.
  • Billing: prepaid credits with a double-entry ledger, per-key spend caps, per-request usage accounting (provider-reported tokens when the upstream supplies them). No inference markup at launch; a 5% fee on credit purchases at the time. (Since 2026-10-29 the card fee is 5.5% + $0.50, added on top; there is no fee on requests.)
  • Analytics: public market analytics — daily usage and the listed-discount distribution across the book.
  • Docs: guides for the OpenAI and Anthropic SDKs, Cursor, Cline, Roo, Aider, Continue, LangChain, and the Vercel AI SDK.