AI Cost CalculatorHow to read these prices
Prices are per million tokens (a token is just under a word), with separate rates for the text you send in, the repeated text you reuse from cache, and the text the AI writes back. The reply is almost always the most expensive of the three, so anything that writes a lot costs more than the input price alone suggests. A sticker price tells you very little on its own - it only means something once you run it against your own mix of in and out.
Full pricing table
| Provider | Model | Input /1M | Cached /1M | Output /1M | Context |
|---|---|---|---|---|---|
| OpenAI | GPT-5.5 | $5 | $0.5 | $30 | 400,000 |
| OpenAI | GPT-5.4 | $2.5 | $0.25 | $15 | 400,000 |
| OpenAI | GPT-5.4 mini | $0.75 | $0.075 | $4.5 | 400,000 |
| OpenAI | GPT-5.4 nano | $0.2 | $0.02 | $1.25 | 400,000 |
| OpenAI | GPT-4o | $2.5 | $1.25 | $10 | 128,000 |
| OpenAI | GPT-4o mini | $0.15 | $0.075 | $0.6 | 128,000 |
| Anthropic | Claude Fable 5 | $10 | $1 | $50 | 1,000,000 |
| Anthropic | Claude Opus 4.8 | $5 | $0.5 | $25 | 1,000,000 |
| Anthropic | Claude Sonnet 4.6 | $3 | $0.3 | $15 | 1,000,000 |
| Anthropic | Claude Haiku 4.5 | $1 | $0.1 | $5 | 200,000 |
| Gemini 3.1 Pro | $2 | $0.2 | $12 | 1,000,000 | |
| Gemini 3.5 Flash | $1.5 | $0.15 | $9 | 1,000,000 | |
| Gemini 2.5 Pro | $1.25 | $0.125 | $10 | 1,000,000 | |
| Gemini 2.5 Flash | $0.3 | $0.03 | $2.5 | 1,000,000 | |
| Gemini 2.5 Flash-Lite | $0.1 | $0.01 | $0.4 | 1,000,000 | |
| xAI | Grok 4.3 | $1.25 | N/A | $2.5 | 1,000,000 |
| Mistral | Mistral Medium 3.5 | $1.5 | N/A | $7.5 | 128,000 |
| Mistral | Mistral Large 3 | $0.5 | N/A | $1.5 | 128,000 |
| Mistral | Mistral Small 4 | $0.15 | N/A | $0.6 | 128,000 |
| DeepSeek | DeepSeek-V4-Flash | $0.14 | $0.014 | $0.28 | 1,000,000 |
| DeepSeek | DeepSeek-V4-Pro | $0.44 | $0.044 | $0.87 | 1,000,000 |
| Cohere | Command R+ | $2.5 | N/A | $10 | 128,000 |
| Cohere | Command R | $0.5 | N/A | $1.5 | 128,000 |
| Groq | Llama 3.3 70B (Groq) | $0.59 | N/A | $0.79 | 128,000 |
| Groq | Llama 3.1 8B (Groq) | $0.05 | N/A | $0.08 | 128,000 |
| Together | Llama 3.3 70B (Together) | $0.88 | N/A | $0.88 | 128,000 |
Search (embedding) model pricing
| Provider | Model | Price /1M | Dimensions |
|---|---|---|---|
| OpenAI | text-embedding-3-small | $0.02 | 1,536 |
| OpenAI | text-embedding-3-large | $0.13 | 3,072 |
| Gemini Embedding | $0.15 | 768 | |
| Voyage | voyage-3 | $0.06 | 1,024 |
| Voyage | voyage-3-lite | $0.02 | 512 |
| Cohere | Cohere embed v3 | $0.1 | 1,024 |
| Mistral | mistral-embed | $0.1 | 1,024 |
Price isn't the whole story
How much text a model can hold at once, caching, batch discounts, image or audio support, speed and answer quality all feed into the choice. The model you actually want is the cheapest one that passes your own quality test - which often isn't the cheapest row on the table. Run your numbers through the API cost calculator to rank models by what they'd really cost you.
Every price here comes straight from the provider's official pricing page, verified on the date shown. The data sources page lists each source and the last verification date.
Turn the table into a real estimate
Say one request sends about 1,500 words in and gets 225 back. Multiply each side by its own rate - don't add the rates together or price the whole thing at the cheaper input rate. Then apply the cached rate only to the text that actually repeats.
Do it again with a short answer and a long one. The reply is usually the expensive side, so a product that writes articles can rank the models completely differently from one that just sorts support tickets.
Why the cheapest row often isn't cheapest in real life
The sticker price doesn't include bad answers, retries, the longer prompts a weaker model needs, or the time spent fixing its mistakes. Compare cost per good answer instead. A slightly pricier model is often the cheaper choice when it gets things right the first time.
Use the tiers as a starting point
Budget models are great for sorting, fact-pulling, tagging and routine chat. Mid-tier models balance reasoning and cost for most core features. Premium models earn their place where a better answer clearly pays off. Treat these as starting guesses, not rules - build a small test set and let the models earn their spot. Plenty of products use more than one: a cheap model for the everyday work and a stronger one for the hard or high-value cases.
Mind the fine print around the rate
Check that the listed price applies to your region and your account type. Batch, caching, long-context and priority processing can all have their own rates, and tax and enterprise deals usually sit outside the public list price.
Write down why you chose
When you pick a model, save the test result, the usage you measured, the price you saw and the date. That makes the next review far easier. Come back to it when your workload changes, the provider updates the model, or a new option clearly beats it on quality or cost.
For a quick calculation, open the API cost calculator. For the source links and update dates, see the data sources page.