AI model API cost comparison (Claude, GPT, Gemini)

Choosing a language model by reputation, without looking at cost, can multiply your application’s bill by ten or a hundred: the gap between a small and a large model is often two orders of magnitude. But the cheapest model is not always good enough: what matters is the cost per task solved.

Enter how many requests you make, how many tokens you send and receive on each and, if you like, what share of the input goes to the cache and what discount you apply. You will see the cost with your model, its price rank among Anthropic, OpenAI and Google models, and a table with all of them from cheapest to dearest. Everything runs in your browser.

Providers charge per million tokens, with one price for input and a higher one for output. The cost of a request is: input tokens × input price + output tokens × output price, all divided by a million.

If you use prompt caching, the part of the input read from the cache is billed at a much lower rate (often a tenth or less). An extra discount (for example the batch API’s) applies to the total. The cost for the period is one request’s cost times the requests a day times the days.

What you pay for in a language API

You pay for tokens, not requests. A token is a chunk of text (in English, about four characters). What you send (input) and what the model generates (output) are charged separately, and output usually costs four to six times more.

How to cut the cost

• **Pick the smallest model that does the job well**: the price gap between small and large models is huge.
• **Use prompt caching** if you repeat long instructions or documents.
• **Use the batch API** for work that does not need an immediate answer (it often gives a 50% discount).
• **Limit the output**: ask for concise answers and cap the maximum tokens.
• **Trim the context**: do not send the whole history if you do not need to.

Price is not the same as cost per task

A cheap model that fails half the time and forces retries can end up dearer than a better one. Measure the cost per task solved with your own examples before deciding.

Tokens are not the same across providers

Each provider has its own tokenizer: the same text gives a different number of tokens depending on the model (in recent versions of some models, up to 30% more than earlier ones). If you compare models from different providers, measure real tokens with each one’s counter.

Common mistakes

• Comparing models only on input price, forgetting that output costs more.
• Forgetting the “reasoning” tokens of models that think before replying, which are billed as output.
• Not counting the conversation history, which is resent on every call.
• Trusting old prices: they change often.

Limits of this calculator

• It uses each model’s standard price (no batch or volume discounts, apart from what you enter) and, for Google models with tiered pricing, the up-to-200,000-token tier.
• Google’s cache price is not listed, so no cache saving is applied to those models; cache-write cost and storage are not included.
• It does not include separately charged tools (web search, code execution), images, audio or video, or fine-tuning.
• OpenAI and Google prices were taken from their pages with a tool that summarises the content: check them on the official page before deciding.
• Prices change: the revision date appears next to the result.

Notice

This is an estimate. Rates are set by each provider and can change without notice; confirm the current price before budgeting.