Mistral Large 4: what it is and what it costs

Mistral Large 4 is the new flagship model from Mistral AI, the Paris lab, and it went live this morning as a paid API preview anyone can use. The model weights, which let you download and run it yourself, are promised for the end of October, so today you rent it by the token and later this month you can host it.

The company's launch post on X, from @MistralAI, calls it "Le Chonk" and lists the basics: 1 trillion parameters, 49 billion active at a time, and native support for images as well as text. Mistral's claim is that it is "the best open weights model from US or Europe on aggregated benchmarks." Note the qualifier. That leaves out the Chinese open models, which have led this category for most of the year.

What Mistral says it does

Mistral's chief scientist, Guillaume Lample, opened his launch thread on X (@GuillaumeLample) with a detail that matters more than the parameter count: the reinforcement learning run behind this preview "is still in flight." He says a final version will ship before the end of the month together with the weights. So what you test this week is a work in progress, and the scores should go up.

Mistral's own post points to cyber defense, manufacturing and finance as the jobs the model is strongest at. It also says Large 4 beats closed frontier models on visual grounding, which means locating a specific object inside an image. The model page in Mistral's docs adds a 1 million token context window, function calling, structured outputs and document Q&A. The model ID there is mistral-large-4+1.

Training ran in Mistral's own European data centers, and Mistral sells hosting from Europe on its own cloud. A US shop probably won't care. If you have European clients or data-residency rules, it could settle the question.

The independent numbers so far

Vals AI posted a first outside read within minutes of the launch. @ValsAI reports that Large 4 is "the #1 open-weight model on HLAB," Harvey's benchmark for legal agent work, and #9 among open-weight models on its broader Vals Index.

The Vals model page adds context. On the overall index, counting closed models too, Large 4 ranks 32nd of 44 with about 48% accuracy. Its legal score, around 16%, ranks 6th of 75, which is high for a benchmark that stays hard for every model. Vals also says that finishing its full test run took Large 4 more than an hour and 45 minutes.

My read: a strong open model that leads in a few specialist areas and sits mid-pack overall. If your work looks like the legal or finance tasks it scores well on, test it. For general office work, the closed models still have the edge.

What Mistral Large 4 costs

The sources don't agree on the price yet. Mistral's docs page shows $0.68 per million input tokens, $0.07 for cached input and $2.09 per million output tokens. Vals and the European press report exactly double: $1.36 in and $4.18 out. I can't tell whether the lower figure is a preview discount or an error, so check the price in Mistral's console before you budget against it.

Even the higher figure is below the $2 input and $10 output OpenAI charges for GPT-6.1 Sol. The weights matter more than the price, though. A trillion parameters won't run on an office computer, but cloud hosts will offer it, and a company that needs its data on hardware it controls gets a strong model it can run there.

What this means for a small business

Here is how I'd use it this month:

• If you handle sensitive documents, like legal files, financial records or client contracts, put Large 4 on your shortlist for when the weights ship. Self-hosting is the main reason to choose an open model.
• If you already use ChatGPT, Claude or Gemini for daily work, don't switch because of a launch thread. Wait for the final version and more independent tests.
• If you pay per token for a high-volume task like sorting inbound email, run a small batch through the API next to what you use now and compare the results and the bill.

This is the second open-weight launch today, after Reflection Beam, and each one pushes down the cost of routine automation a little more. For most owners the harder decision is which tasks to hand to a model at all. New Face Design's free process audit works through that with Fox Valley businesses before anyone picks a model.