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

The cheapest GPT API (route it, or switch models in one line)

To use GPT models for the least money you have two levers: route each request to the cheapest healthy seller for the GPT model you want, and switch to a cheaper model when a task does not need GPT at all. MartLLM gives you both through one OpenAI-compatible key — the price shown has no MartLLM fee added, and switching models is a one-line change.

Two levers, not one

Most "cheapest GPT API" advice stops at picking a smaller model. There is a second, bigger lever: the same GPT model is sold by more than one seller, and the live price is whatever the cheapest healthy seller is charging. A gateway routes you to that seller automatically.

MartLLM pulls both levers. You keep the OpenAI SDK and one key, and every GPT request is matched to the cheapest healthy seller. When a task does not actually need GPT, you drop to a much cheaper open-weight model by changing one string. The price on the live pricing board has no MartLLM fee added.

Keep your OpenAI code, change the base URL

from openai import OpenAI

client = OpenAI(
    base_url="https://api.martllm.com/v1",   # was: https://api.openai.com/v1
    api_key="ull_your_key_here",
)

resp = client.chat.completions.create(
    model="openai/gpt-5.4-mini",   # or any id from /models
    messages=[{"role": "user", "content": "Summarise this ticket."}],
)

Nothing else in your OpenAI SDK code changes. See the OpenAI SDK doc for streaming, tools and structured outputs, all unchanged.

When to switch models instead of routing

  • Classification, extraction, routing, short rewrites rarely need a frontier GPT model. Open-weight models (DeepSeek, GLM, Qwen) do them for a few cents per million tokens — compare the live numbers on /models.
  • Long, cheap, high-volume jobs are where the per-token gap compounds. Changing model="openai/gpt-5.4-mini" to a cheaper id is the whole migration.
  • Keep GPT where it earns its price — hard reasoning, agentic tool use — and route it to the cheapest healthy seller there.

Failures are free

A GPT call that fails before output starts is retried on the next healthy seller, and a request that returns nothing costs $0 — how pricing works. No monthly fee, no minimum spend.

Keep reading

Live GPT pricing

ModelInput / 1MOutput / 1M
GPT-OSS 20B$0.000018$0.000089
GPT-OSS 120B$0.000069$0.00030
GPT OSS Safeguard 20B$0.000075$0.00030
GPT OSS Safeguard 120B$0.0015$0.0060
GPT-6 Luna Decisions (OpenAI)$0.0240$0.0000
GPT-6 Luna$0.0030$0.0149
GPT OSS 20B$0.0125$0.0475
GPT 5 Nano$0.0097$0.0773
GPT-4o Mini$0.0261$0.1045
GPT-5.6 Luna$0.0200$0.1200
GPT 5.4 Nano$0.0324$0.2027
GPT-6 Sol$0.0414$0.2070

Live prices via MartLLM, no fee added · full pricing table →

FAQ

Is there a cheaper GPT API than OpenAI direct?

MartLLM routes each GPT request to the cheapest healthy seller for that model, and makes it a one-line change to switch cost-sensitive work to much cheaper open-weight models — all through one OpenAI-compatible key, no fee added to the price shown.

Do I have to rewrite my OpenAI SDK code?

No. Change only the base_url and api_key. Streaming, tool calls and structured outputs all work the same. Switch models by changing the model id string.