AI API Cost Calculator

About the AI API Cost Calculator

Work out what an AI feature will cost before you build it. Enter how many requests it makes a day, how many input and output tokens each request has, how much of the input the prompt cache serves and how many requests can go through the Batch API, and pick a model: the calculator shows the cost per request, per day, per month and per year, and ranks the other models for the same workload.

The presets hold the per-million-token prices that OpenAI, Anthropic, Google and Mistral AI publish for input, cached input, cache writes, output and batch requests, with the long-context prices that some models charge for very long prompts. Every price can be edited, and you can add your own model — a fine-tuned one, a negotiated rate or another provider. Paste a sample prompt and answer to count their tokens with the same engine as the LLM Token Counter.

Common uses

• Checking that an AI feature pays for itself before you write the code, or before a pricing decision.
• Choosing between a large and a small model, or between providers, for the same workload.
• Seeing how much prompt caching and the Batch API save on a high-volume job.
• Setting a monthly API budget and knowing how many requests it covers.

How the cost is worked out

Each token type has its own price per 1M tokens:

• Cost per request = (input tokens not cached × input price + cached tokens × cached-input price + (output tokens + reasoning tokens) × output price) ÷ 1,000,000.
• Requests sent through the Batch API use the batch prices; the share you enter is blended in.
• A cache write replaces a cache read for the cached part of that request, at the cache-write price where the provider lists one and at the input price where it does not.
• Per day = cost per request × requests per day + cache writes. Per year = per day × 365, and per month is a twelfth of the year (30.42 days).

Prompt caching

When many requests start with the same text — a long system prompt, a document, tool definitions — the provider can serve that part from its cache at a much lower price. The cached share is the part of the input tokens read from the cache. Anthropic charges extra to write the cache (1.25 times the input price for the 5-minute cache, 2 times for the 1-hour cache) and charges 0.1 times the input price to read it, less on some models, as its pricing page lists; some OpenAI models also list a cache-write price on the OpenAI pricing page. A request that finds no cache pays for writing it, so enter roughly how many times a day the cache is created.

Batch requests and long prompts

The Batch API returns results later instead of at once, at lower prices: Anthropic’s page states a 50% discount on input and output tokens, OpenAI and Google list batch prices on their pricing pages, and Mistral’s batch documentation says batch workloads run at a 50% discount. Some models charge more for very long prompts: OpenAI’s long-context prices apply to the whole request above 272,000 input tokens, and Gemini Pro models charge more above 200,000 tokens (Gemini API pricing). The calculator applies these automatically and says so in the notes.

Reasoning and thinking tokens

Reasoning models think before they answer, and those tokens cost money even though you never see them: OpenAI bills reasoning tokens as output tokens (reasoning guide), Google’s output prices include thinking tokens, and Anthropic reports thinking as part of the billed output tokens (extended thinking). Take the average from the usage figures of a few test calls: OpenAI reports output_tokens_details.reasoning_tokens and Anthropic output_tokens_details.thinking_tokens, and in both the output_tokens total already includes them. Enter the rest as output tokens and the reasoning separately, or the whole total as output with no reasoning tokens, but do not count them twice.

Where the prices come from

The presets are copies of the per-million-token prices on the providers’ own pages: OpenAI, Anthropic, Google Gemini API and Mistral AI, in US dollars. Prices that the provider lists as changing later, and sale prices, are marked, with a second preset at the later price or the list price. Prices change, so open the provider’s page before you commit to a budget — and edit any price the calculator shows.

Limitations

• The presets are copies of published list prices: check the provider’s page before you rely on them, and enter your own rate if you have a discount or credits.
• Taxes, free tiers, regional or data-residency surcharges, priority or fast processing, tool fees (such as web search per call), images, audio and the per-hour storage of explicit context caches are not included.
• Token counts of a sample are exact only for tokenizers that are published (OpenAI’s and the open models’); Claude and Gemini counts are the providers’ rule-of-thumb estimates.
• Real requests also carry tokens you may not count: chat formatting, tool definitions, retrieved documents and retries. Compare with the usage report of a few test calls.
• The calculator does not check a model’s context window: make sure your prompt and answer fit the model you choose.

Privacy

Everything happens in your browser. What you enter or open here is not uploaded or stored by MySmartCoPilot. Counting a sample prompt downloads the tokenizer you pick once from MySmartCoPilot; your text and numbers never leave the browser.