How we verify LLM API pricing

How we price LLM APIs

Four rules, the same for every model.

1. One source of truth per provider. Every per-token rate comes from the vendor's own official pricing page (listed below) — not a blog, aggregator or screenshot.
2. Verified by hand, then promoted. A scrape or research pass is only evidence; a number reaches our data file only after a human confirms it against the official page. Raw dumps are kept dated for audit.
3. Computed, not quoted. The "≈ $/mo" you see is calculated by our engine from input, output and cached rates at a stated workload — so models compare on one number. We never invent a monthly price.
4. Every change is logged. Prices are committed before the site rebuilds, so each change is dated in the public changelog with the provider it came from.

Sources — every number traces to an official page

28 models across 6 providers. Last full verification: 2026-06-20.

Mistral's per-model lineup is taken from the official model cards on docs.mistral.ai; the public FAQ price page lists only the flagship.

Are tokens the same across models? No — and it matters

A "token" is not a fixed unit. Each model family uses its own tokenizer, so the same text becomes a different number of tokens from one provider to the next — commonly a 10–30% spread (wider for non-English text and code). Two models at the same "$ per 1M tokens" can therefore cost differently for the same prompt, because one encodes it in fewer tokens.

Tokens aren't equal in value either: a token from a frontier model carries more capability than one from a budget model, so price per token says nothing about quality.

What this means for our numbers: we compare list price per 1M tokens as the billing unit — the honest, vendor-published basis — and run every model through one identical workload (same token counts) so the ≈ $/mo column is apples-to-apples on price. Treat cross-provider gaps inside ~10–20% as a tie until you test your own prompts; the calculator lets you plug in your real token mix.

What each number means

Known limits & confidence

We would rather show a gap than a guess:

• Mistral cached prices are held empty (except the flagship): Mistral documents a caching rule but doesn't publish per-model cached rates, so we bill input at the full rate rather than estimate.
• Batch multipliers for xAI and DeepSeek are held empty: a batch or cache discount exists, but the exact figure isn't officially tabled, so we don't apply one.
• Long-context surcharges (e.g. Gemini above 200k tokens) are noted in our data but not yet in the calculation — short-context pricing is shown.
• Tier and our example workload are assumptions, not vendor facts — change them in the calculator for your case.

What we exclude, and why

• Cohere — the current lineup is sold as enterprise instances with no public per-token price, so it can't be verified the same way.
• Meta Llama — priced only through third-party hosts with their own rates; planned for a later pass once we can compare hosts fairly.

We track the 28 models we can verify against an official, first-party price. We would rather cover fewer models accurately than more on guesswork.

How we keep it current

An automated job re-reads all 6 official price pages every morning and flags any change; a second job watches context windows, caching, batch and deprecations. Those flags are evidence — a human still confirms before a number changes. Confirmed changes are dated in the changelog, and we take no commission from any model provider, so nothing here is pay-to-rank.

Frequently asked questions