Comparison

Understanding what it costs to run AI models.

AI API prices change often, and every provider presents them its own way. We gathered them in one table: input, output and cache prices for 20 models from OpenAI, Anthropic, Google, Mistral and Scaleway, collected on 5 August 2026. Every row carries a link to the official price list and an indication of where your data is processed.

Prices checked on 5 August 2026 against the providers' official pages. Updated monthly.

A token

A token is a run of characters: roughly four in English. Providers bill by volume of text, never by number of user requests: a thousand short questions cost less than a hundred questions with a long document attached.

Input

Everything you send the model: the question, but also the instructions and any attached documents. It grows fast as soon as you add a conversation history or a file.

Output

What the model produces. It costs more than input on all 20 models in the table: from 3 to 8 times more.

Input cache

Send the same text twice, a document or a conversation history, and some providers bill the second time far less than the first: that is the input cache. It matters most on long-context conversations, where the history grows every turn and gets sent again in full with each message. The table shows it in the "Input cache" column, when the provider publishes one.

The table.

The context window is not in this table, because it is not billed the same way from one provider to the next. At Google, going past 200,000 context tokens moves you to a higher rate. At Anthropic, a long context stays billed at the standard price.

Simulator

Find the cheapest model for your usage.

Which model is cheapest depends on your usage. An assistant that goes back and forth over the same context benefits from the input cache: at some providers, that can divide the price by ten. An agent that produces long answers, on the other hand, mostly pays the output price.

Reasoning models

The hidden costs behind the price per token.

The number of requests does not move, the invoice does. Three mechanisms change what you actually pay, and none of them shows up in the price per million tokens.

Reasoning tokens

A reasoning model produces intermediate tokens before its answer. They are billed at the output rate and do not appear in what the user reads. Their volume is a setting: Anthropic requires at least 1,024 tokens and suggests starting at 16,000 on a complex task, OpenAI states "a few hundred to tens of thousands" depending on difficulty. On a 400-token answer, that means 3 to 40 times more billed tokens than the user ever sees.

The input cache

An option, available only at some providers, that can divide by ten the price of the reused part of a long conversation. Mistral and Scaleway do not publish one: with them, the conversation history is billed at the full input price on every request.

Batch processing

A third lever on the invoice: batch processing, which generally offers a half-price discount on anything that can wait a few hours rather than answer immediately.

Data residency

Three questions, not a checkbox.

"Hosted in Europe" means nothing until three questions are answered separately: where inference runs, where data is stored at rest, and under which contract. The table column answers yes or no, and the detailed answer is here, provider by provider. Read the scope above all: at Google, the option exists on a product whose prices are not the ones shown.

OpenAI: yes

Europe residency (EEA and Switzerland) on the eu.api.openai.com endpoint, for processing as well as storage. It is configured project by project, and requires two steps: signing an amendment on data retention, and getting OpenAI's approval on its abuse-monitoring controls. Watch out if you use the Assistants API: the files and threads it stores are not covered by this residency guarantee, even if the rest of your usage is. Provider documentation.

Anthropic: no

The Anthropic API offers only two inference geographies, "global" and "us", and storage at rest is available in the United States only. EU residency remains possible through Amazon Bedrock or Google Cloud in European regions: the contract and price list are then those of the partner cloud, not the ones shown here. Provider documentation.

Google: yes

Gemini Enterprise lets you pick an "eu" multi-region, which covers storage at rest as well as processing, with per-country regions including France for some models. Mind the scope: the prices listed here are those of the Gemini developer API, whose documentation describes no residency option at all. Gemini Enterprise is a different product, with its own price list. Provider documentation.

Mistral: yes

European infrastructure, with a regional endpoint api.eu.mistral.ai that serves only the models hosted in that region and guarantees that the data it processes does not leave it. A US endpoint also exists, but you have to pick it explicitly. Provider documentation.

Scaleway: yes

Models served from the Paris region, and prices published directly in euros, which removes the currency risk from the budget line. The catalogue is made of open models: you will not find the proprietary models from OpenAI, Anthropic or Google. Provider documentation.

From price list to invoice

Take back control of your AI budget.

An API invoice tells you how much you paid. It does not tell you which team, which project, or which model. optidome makes that connection, whether the spend is billed per token, per licence or in prepaid credits.

Frequently asked questions

How do I read a price per million tokens?

Providers bill per token, separately for what you send (input) and what the model produces (output). A token is roughly four characters. So $2 per million input tokens means a million characters sent costs about $0.50.

Why are some prices in dollars and others in euros?

Because each row keeps the currency the provider publishes in. OpenAI, Anthropic, Google and Mistral publish in dollars, Scaleway in euros. Converting at rest would freeze an exchange rate into the data; the conversion happens at display time, at the European Central Bank reference rate whose date is shown.

Why do reasoning models cost more at the same number of requests?

Because they produce intermediate tokens, billed at the output rate but absent from the answer on screen. At a constant number of requests and an identical published price, the same task can cost several times more depending on the model. Since those tokens appear neither in the question nor in the answer, nothing signals them before the invoice.

What does the EU residency column mean?

It says whether the provider itself offers an option to process and store data in the European Union. A "no" means the option exists only through a partner cloud, with a different contract and a different price list. A "yes" does not always apply to the model on that row: the per-provider detail higher up the page gives the exact scope, because inference location, storage at rest and contractual terms do not overlap.

How often is this comparison updated?

Every month, on the official pricing pages. The date of the last check is shown at the top of the page and every row links to its provider price list.