Get started
Text model

Mistral — compact models from the European lab

Mistral models in GPTunneL: open weights, the compact Ministral tier and multimodal Pixtral. Pay per token, no subscription and no VPN.

What the Mistral family is good at

Traits Mistral AI keeps from one generation to the next — you can count on them whichever model you pick.

European engineering

A French company with European data centres — a decisive argument wherever the provider's jurisdiction is a requirement.

Genuinely compact models

Ministral at 3 and 8 billion parameters runs fast and costs next to nothing: it is the tier for a stream of repetitive requests.

Open weights

A large part of the family is published under a free licence and can run on your own server.

A dedicated branch for images

The Pixtral models understand images — they read diagrams, screenshots and scans together with the text of your request.

Which model to pick

The family has three tiers — for different jobs and budgets.

MinistralVolume and speedMediumThe default pickLargeHard tasks
Best forClassification, tagging, simple extractionEveryday writing, correspondence, analysisHard reasoning and large documents
Contextup to 131K tokens131K tokensup to 262K tokens
SpeedHighestHighMedium
PriceLowestLowAbove average

Figures come from the platform catalogue. Exact per-model prices are on the Pricing

Every Mistral model

Release dates follow Mistral AI's official announcements.

December 2025

Mistral Large 2512

The current top model: context grew to 262 thousand tokens, with image understanding and web search added.

May 2025

Mistral Medium 3

The middle tier that covers most everyday tasks at a noticeably lower price than Large.

November 2024

Pixtral LargePixtral 12B

The multimodal branch of the family — it understands images on a par with text.

October 2024

Ministral 3BMinistral 8B

Very compact models at 3 and 8 billion parameters — built for fast, repetitive requests.

July 2024

Mistral Large

The generation that brought the company up to the level of the top closed models.

Full release history
April 2024

Mixtral 8x22B

An open mixture-of-experts model: the quality of a large model at the compute cost of a small one.

February 2024

Mistral Small

The company's first commercial models, and the start of the current family.

How billing works

There is no subscription: you top up one balance and spend it on any model on the platform.

Pay per token

You are charged for exactly what the request and the answer used. Not using it costs nothing: no limits, no monthly fee.

One balance for every model

Mistral, ChatGPT, Claude, image and video generation — all out of the same wallet. No separate subscription per service.

Top up the way you prefer

International cards, Apple Pay, Google Pay or crypto. The minimum top-up is $5.

See pricing

How to get started with Mistral

Sign in to GPTunneL
One account for every model on the platform.
Cut this email down to three sentences, keeping the ask and the deadline
Mistral Large
Shortened email

We are asking you to approve the phase two estimate — it is up 12% on currency movement.

Phase one is complete and the handover documents are with you for signature.

We need an answer by Friday, otherwise the installation start slips.

Sign in whichever way suits you.

The model switches right inside the input — you can change it mid-conversation.

The answer lands in the chat and the context is kept until the conversation ends.

Try Mistral in GPTunneL

Signing up takes a minute, and $5 on the balance is enough to see whether the model fits your task.

Frequently asked questions

No. Requests go through GPTunneL's infrastructure, so the models open like any ordinary website — no VPN and no proxy.

Mistral is a family of neural networks from the French company Mistral AI, the leading European player on the market. The models answer questions, write and analyse text and program, while a separate Pixtral branch understands images.

The family is chosen for two reasons. The first is jurisdiction: the company and its data centres are European, and where the provider is subject to requirements that settles the question. The second is compactness: the Ministral models at 3 and 8 billion parameters run very fast and cost next to nothing, which makes them the choice for a stream of repetitive requests. A large part of the family is published under a free licence and can run on your own server.

Mistral models are available in GPTunneL — worldwide, without a VPN, with one balance and billing for the tokens you actually spend.