LLM API Pricing Index

As of 2026-10-11, across 451 models from 61 vendors: the median paid model costs $0.86 per million tokens blended (three input to one output), the cheapest paid model is Mistral: Mistral Nemo at $0.029, and the frontier — Anthropic: Claude Fable 5.1 $20.00, OpenAI: GPT-6 Astra $20.00, SpaceXAI: Grok 4.6 $3.00. 19 models are free-tier.

Three input tokens for every output token, which is the ratio a typical chat or RAG workload runs at: blended = (3 × input price + 1 × output price) ÷ 4. Output tokens are usually three to five times dearer than input, so a model's blended price says more about what you will pay than either figure alone. Both are shown.

OpenRouter's public model list, which publishes the per-token price it charges for each of 451 models — for nearly all of them the vendor's own list price passed through. It is the one place every vendor appears in one schema. A vendor's direct price can differ, especially with volume discounts, cached input or provisioned throughput; the vendor's own pricing page is linked on every row.

Vendors cut prices often and raise them rarely. The list is read once a day and every snapshot is kept, so a model's 7-day change is shown where it has one, and the history is in the repository. Nothing is estimated.

Renting a GPU and running an open model yourself is the alternative to paying per token. The GPU index prices that: each GPU page estimates cost per million generated tokens for open models at today's rental median, so the two can be compared for the same job. At low volume the API wins; at sustained load a rented card usually does.