AI model pricing calculator

As a rule of thumb, a token is about four characters or three quarters of an English word, so 1,000 tokens is roughly 750 words; other languages and code usually take more tokens per word, and every model family tokenizes a little differently. When you set tokens per request, count everything the model reads as input: the system prompt, retrieved documents, tool schemas and the conversation so far. Reasoning models also bill their thinking as output tokens. For exact figures, FastRouter's request logs report the token counts of every call.

Input and output tokens are priced separately per million tokens, and output is usually several times dearer because generating text costs more compute than reading it. The estimate multiplies your requests per month by the tokens per request on each side, applies the model's cached-input rate to the share of input you mark as cached, and adds the two together. Batch and Flex tiers swap in that tier's rates where the model offers them.

Image models charge per image, usually by size and quality. Video models charge per clip at a given duration and resolution, or per second of output, with a separate rate when audio is generated too. Audio models charge per second of audio, per character of text, or per audio token. The calculator exposes exactly the options the catalog prices for each model, so you only see choices that change the bill.

It covers the metered rates shown for each model. Long-context surcharges, web search, cache writes, tool fees and provider-specific pricing are not included, and providers change their rates over time, so treat the result as an estimate rather than a quote. Your invoice follows actual usage at the rates in force when each request runs.

Four FastRouter features lower the bill at the same catalog rates, and text models in the calculator have a control for each. Prompt caching bills repeated context at the model's cached-input rate, as low as 0.10× the input price on supported providers. Prompt compression trims input tokens before they reach the provider, with the default engine keeping about 75% of them. Response caching returns identical or similar requests at 0.1× standard pricing with free storage. Flex pricing routes requests that can wait to the provider's discounted tier. Insights then audits your real traffic weekly and shows where those savings actually are.

The typical mix is a conservative starting point, not a prediction: 20% of input tokens read from the prompt cache, 10% fewer input tokens from compression, and 10% of requests answered from the response cache. Figures reported for favourable workloads run far higher, prefix-cache hit rates of 70 to 90% on stable system prompts, 2 to 5 times compression on long retrieved context, and 40 to 70% response-cache hits on FAQ-style traffic, but general chat sees much less, so the example stays well below them. Prompt caching also needs a repeated prefix of at least 1,024 tokens, so the example skips it for shorter prompts. The arithmetic runs in order: compression removes its share of input tokens, prompt caching reprices the cached share of what remains at the model's cached-input rate, then response-cache hits reprice that share of requests at 0.1×. Every share is editable, and Insights reports the real figures for your traffic.

Yes. Add as many models as you like across types and the summary totals them, so you can price a whole pipeline, for example a chat model plus image generation plus text-to-speech. Adding two text models with the same usage is also the quickest way to compare their cost side by side; the compare pages add benchmarks, context windows and per-provider prices on top.

No. Model usage is billed at the provider's list price with $0 markup and no commission, so the rate the calculator shows for a model is what that model costs through the gateway. FastRouter earns on the platform around the models, not on the tokens, images or seconds you generate.