Calculator
What will it cost?
7 models, priced against your workload. The cheapest one changes depending on how much you read versus how much you write.
Retrieval stuffs the context. Input dominates, so input price decides the winner.
| # | Model | Vendor | Input | Output | Per month | vs cheapest |
|---|---|---|---|---|---|---|
| 1 | Llama 4 Freefree | meta | $0 | $0 | $0 | — |
| 2 | Gpt 5 Mini | openai | $7.50 | $1.80 | $9.30 | 1× |
| 3 | Deepseek V4 | deepseek | $13.50 | $3.30 | $16.80 | 1.8× |
| 4 | Gemini 3 Pro | $62.50 | $30 | $92.50 | 9.9× | |
| 5 | Claude Sonnet 5 | anthropic | $150 | $45 | $195 | 21× |
| 6 | Gpt 6 Astra | openai | $200 | $60 | $260 | 28× |
| 7 | Claude Opus 5 | anthropic | $250 | $75 | $325 | 34.9× |
No model matches those constraints. Try a smaller minimum context.
List prices for pay-as-you-go API access, excluding volume discounts and batch rates. Cached-input savings are applied only where the provider publishes a cache-read price. See methodology and the full model table.
Enter what you expect to send and receive in a month and this ranks every model by what it would actually cost you. Prices come from the same daily feed as the rest of the site.
Questions
Which model is cheapest?
It depends on your workload, and that is the point of this page. A model with cheap input and expensive output wins on document Q&A, where retrieval fills the context, and loses on content generation, where the model writes far more than it reads. Switch between the presets above and watch the order change.
What is a token, roughly?
About three quarters of an English word. A thousand tokens is roughly 750 words, or a page and a half. If you have no idea what your volume is, start with a preset and adjust — the ranking is usually more stable than the absolute number.
What does cached input do to the bill?
Several providers charge much less for input they have seen before, which matters when a long system prompt or document is resent on every call. The slider applies that discount only to models whose provider publishes a cache-read rate. For the rest the discount is simply not applied — we do not invent one.
Are these the prices I will actually pay?
They are the published list prices for pay-as-you-go API access, before volume discounts, committed-spend agreements or free tiers. Batch or asynchronous processing is often cheaper again. Treat the result as a well-founded estimate, not a quote.
Why is a model I use missing?
Because it has no published per-token price in our source, or only one side of the price is published. A model is left out rather than shown with a guessed number.