How these numbers work
A price comparator is only worth reading if you know what it is measuring and where it stops being precise. This page is that boundary, written out.
Where the prices come from
Rates are pulled from the LiteLLM pricing dataset, an open, community-maintained record of published provider rates, and normalised into USD per million tokens. The sync runs as a script, not by hand, so a stale figure is a missed cron job rather than a forgotten edit. Every page carries the sync date; the build fails if the data is more than 30 days old.
Currently 246 models across 12 providers. Only first-party rates are listed. Resellers and inference platforms — Bedrock, Azure, OpenRouter, Fireworks, Together — often price the same model differently, and mixing them into one table makes it look like a model has several prices.
Which token counts are exact
45 models are counted exactly. OpenAI publishes its tokenizer, so your text is run through the real encoder — the same segmentation the API bills against.
Every other family is estimated. Anthropic, Google, Mistral, xAI, Moonshot and the rest do not publish a tokenizer, so the count is the OpenAI count scaled by a published ratio for that family. Expect roughly ±10%, and more on code or non-English text, where tokenizers diverge most.
This distinction is on every model, as an exact or estimated badge. Several comparable tools apply the OpenAI tokenizer to all models without saying so, which reads as precision it does not have.
Long-context rates
28 models charge a higher rate once a single request passes a context threshold. When your prompt crosses it, the higher rate is applied automatically and the row is flagged — you do not have to know the threshold exists to avoid being misled by the base figure.
| Model | Base in | Threshold | Above |
|---|---|---|---|
| gpt-5.6-luna | $0.20 | 272K | $0.40 |
| grok-4-1-fast | $0.20 | 128K | $0.40 |
| grok-4-1-fast-non-reasoning | $0.20 | 128K | $0.40 |
| grok-4-1-fast-non-reasoning-latest | $0.20 | 128K | $0.40 |
| grok-4-1-fast-reasoning | $0.20 | 128K | $0.40 |
| grok-4-1-fast-reasoning-latest | $0.20 | 128K | $0.40 |
| grok-4-fast-non-reasoning | $0.20 | 128K | $0.40 |
| grok-4-fast-reasoning | $0.20 | 128K | $0.40 |
| gemini-pro-latest | $1.25 | 200K | $2.50 |
| gemini-2.5-computer-use-preview-10-2025 | $1.25 | 200K | $2.50 |
What the cost estimate excludes
- Batch and off-peak discounts
- Negotiated or enterprise rates
- Free tiers and trial credits
- Reasoning tokens billed beyond the visible output on thinking models
- Tool calls, web search, and other per-request surcharges
- Retries, which are invisible until the invoice arrives
Cache pricing is included where published: read rates for every model that has one, and separate write rates for the 22 models that bill cache creation on its own — Anthropic charges it at two TTLs.
Output length is an input you choose, not a measurement. The presets scale it from the prompt because that is the closest honest proxy; a real workload is the only way to know.
Deprecation
33 listed models have a published deprecation date. They are hidden from the picker by default and marked on provider tables, since pricing one into a plan you will still be running next year is a trap.
What we would get wrong first
Provider pages change without notice, and the upstream dataset can lag a price change by days. The counts for non-OpenAI families are the softest number here — if a decision turns on a difference of a few percent, count your real payload against the provider's own token-counting endpoint instead of trusting this page.