NewHow we catch a seller listing the wrong model, and refund you if one does.

Pricing

This page describes how requests are priced and billed.

How pricing works

Each request is sent to an available third-party provider for the model. Any provider that is up can serve it. When nobody can serve the model right now, the request answers 503 no_available_provider and nothing is charged.

A request costs what its provider charged for it. No MartLLM fee on requests. Adding credit by card costs 5.5% + $0.50.

Who the providers are and why prices are often below list: How MartLLM works.

"Price now" and "% off" on martllm.com and on each model page are the marketplace's own numbers, refreshed every 30 seconds: the lowest price a provider is selling the model at right now, and how far that is below the model maker's list price. A request is often served at that price, and can cost more when the cheapest providers are busy.

The listed price is what a request typically costs: pricing from GET /v1/models and price_now from GET https://api.martllm.com/api/v1/public/models. It is not a ceiling: a call served by a pricier provider costs what that provider charged. The listed price is the rate a request is held at before it runs, and the rate it is billed at when the provider does not report a cost. It is updated every 5 minutes: the provider's list price times the share of the list price the marketplace actually charged in its most expensive traded hour of the last 24 hours (never lower than the cheapest offer). When nothing traded in the last 24 hours, it is the list price. When no provider sells the model at that price, the listed price is the cheapest provider's real price, which can be above the list price; it drops back by itself when a cheaper provider returns. A model that no route can serve right now has no price now; only its list price is shown.

What one call costs. When the provider reports what it charged for a call, the call costs exactly that amount. When it does not, the call costs its tokens at the listed price (cached prompt tokens at the listed cache-read price, prompt tokens written to the cache at the listed cache-write price, pricing.cache_write_usd in GET /v1/models; a model with no cache-write price lists them at its input price). A call often costs less than the listed price, for example when part of a prompt was read from cache. A call the provider charged $0 for (providers bill in whole millionths of a dollar, so a very short call can cost them nothing) costs $0. The cost of every call is listed in Dashboard → Activity. A request that fails is not billed.

Listed and official cost of each call. Activity (and its CSV) shows, next to what a chat call cost, what the same tokens cost at the listed price when the call was made, what they cost at the official price (the model maker's own price for that call), and how far below the official price you paid. A call that cost more than the listed price is marked, with the reason: on Balanced routing that is a faster provider; set the key to Lowest price to always pay the lowest price. Providers reprice minute to minute, so the listed price of a call is the one when it was made, which can differ from the price on the pricing page now. Calls before 2026-11-10 do not have these figures.

Want a hard limit? Turn on Never charge more than the listed price for your key on Dashboard → API keys, or set it to Lowest price (routing settings). Or send provider.max_price on a chat request: it is only sent to providers at or under that price, and it is never billed above it. For images, video and music, send max_cost: a request that can cost more is refused, and a job is never billed above what is held for it (max_cost_usd).

A seller's own offer (a route marked "seller") is billed at that offer's listed price. Every seller offer is checked to serve the model it claims before it gets any requests, and keeps being checked on real traffic. A seller caught serving a different model is banned, and the buyers it served are refunded from its unpaid earnings. If a seller reports far more prompt tokens than your request holds, you are billed for at most about 2× our own count.

% off list is the marketplace's own figure: its best live price against the model maker's list price. No % is shown under 1%.

Official price: one source, the same on every page. When the marketplace gives no list price above ours, the official price shown is, in this order: the model maker's own published price (for example Google, OpenAI, Kling, BytePlus), else OpenRouter's (read every hour), else Fal's. The first one that prices the model in the same unit we do is the official price, labelled "Official price · ". It is shown only where it is above our price; a cheaper source further down the list never stands in for it. Every maker and Fal price is kept with the page it came from and that page's own words, and the page is read again every day: if the words are gone (the price changed), the price is hidden until it is checked again. Video is priced at the lowest resolution we offer; images at the default size. The homepage, the model board, model pages, comparison pages and GET /v1/models (list_usd / list_input_usd + list_source) all show this same price. In GET /api/v1/public/models it is media_ref (usd, source, url, checked) on media models, and list_input_usd_per_mtok / list_output_usd_per_mtok with list_source on chat models. A web-search (:web) or encrypted (e2ee-) version of a chat model shares its base model's official price.

Fal video follows the cheapest Fal seller. Fal video is priced at Fal's own price less the discount of the cheapest Fal seller that is live, has room left today and has run one of our jobs in the last 24 hours (the seller list is read every 5 minutes). A job is billed at the seller that ran it, never above the price shown.

Images, video and music are priced per unit (an image, a megapixel, a second of video, a job, a minute). The price now is worked out the same way, per unit. A job costs what the provider charged for it, and never more than what is held for it while it runs (max_cost_usd, which stays within a max_cost you send). When the provider never reports a charge, the job is billed at the listed price of what you asked for (the cost_usd of its quote: the price now times the images, seconds at the resolution asked, or other units). Fal video is priced when you submit it, at the price its quote showed. See Make a video or image.

Prepaid credits + card fee

Inference is prepaid. New accounts start at $0. Add credits before your first call. You pick the credit you want; a card fee of 5.5% + $0.50 is added on top when you pay, so your balance goes up by exactly the amount you picked. This is the only fee MartLLM charges; requests carry no fee. Examples:

Credit you pick Card fee You pay
$5.00 $0.78 $5.78
$10.00 $1.05 $11.05
$50.00 $3.25 $53.25
$100.00 $6.00 $106.00

(You pay = credit × 1.055 + $0.50, rounded up to the cent.) The fee covers what the card processor charges us. Credit is between $5 and $10,000 per top-up, paid by card or PayPal on Dashboard → Credits (the fee is the same for both), which shows the fee before you pay; your balance is drawn down per request at the price each call actually cost.

Spend controls

  • Per-key spend limits — set a daily / monthly / lifetime cap when you create a key, or change it any time. Requests that would exceed it return 402 key_spend_cap_exceeded.
  • Per-key rate limits — an optional requests-per-minute ceiling per key.
  • See it per key — Dashboard → Activity filters by key, model and result, totals spend per key, and downloads what it shows as CSV.
  • Balance floor — when credits run out, requests return 402 insufficient_credits until you top up. Reservations for in-flight requests are released if the request aborts or fails. A chat request's reservation covers its whole prompt: text, tool definitions, tool calls and results, response schemas, images, audio and files. An image sent as a link is reserved at the most that model bills for one image, since its size is not known until the provider opens it. You are billed for what the provider counted, and the rest is released.

See Errors for the exact 402 shapes.

What you saved, and sharing it

Once your calls have saved at least a cent against the model makers' list prices, the dashboard shows a green line on Overview, Activity, Credits and Affiliate Program: "You saved $210.40 vs list price in the last 30 days". Nothing is shown before that; there is never a card of zeros.

How the number is worked out, per successful call:

  • list price = the same tokens at the model maker's official price (cached input at the official cache-read price);
  • a call never counts for less than it cost you, so a call that cost more than list adds no negative saving;
  • pictures, video and music have no token list price and count at cost: no saving is claimed for them;
  • calls made with your own provider key (BYOK) and failed calls are left out;
  • percentages are rounded down.

Share makes a card of those numbers in one click: a link (https://martllm.com/s/<id>) and a picture you can post on X or LinkedIn, or copy. A card is a snapshot. It does not change later.

A public page that stays up to date (in the same Share window) is one link that always shows your last 30 days. It is off until you turn it on. Turn it off and the link stops working within 15 minutes; turn it on again and the same link works again. If the last 30 days saved nothing, the link shows nothing.

What a card or page shows: what you paid, the list-price value, the number of calls, the period, up to three model names, and your referral code. It never shows your prompts, answers, API keys, name or email, and nobody can find it without the link. The Privacy Policy says the same.

Dashboard API (signed in): GET /api/v1/usage/savings?days=7|30|90, POST /api/v1/usage/savings/share {"days":30}, and GET / PUT /api/v1/usage/savings/page {"on":true}. Public: GET /api/v1/public/savings/<id> and /api/v1/public/savings/<id>/card.png.