· AI
Mistral's New Model Isn't Open Yet, and It Isn't Cheaper Than China's
France already decided Large 4 is worth paying a premium for. A business spending its own money should wait for the actual numbers.
Mistral unveiled its new flagship, Large 4, on October 6, branding it a sign that Europe can compete with the United States and China on frontier AI. CEO Arthur Mensch told reporters the model sits "above the Chinese models on certain aspects, including cyber." I think Large 4 is a genuinely good model. I also think two claims built into this launch, that it is open and that it beats the Chinese open-weight field, do not hold up yet against the numbers Mistral itself published, let alone the one independent benchmark firm that has tested it so far.
Start with the strongest case for Mistral, because France already made it with real money. In January 2026 the French Ministry of the Armed Forces awarded Mistral a framework agreement covering defense agencies and related institutions, running on sovereign French infrastructure rather than a model hosted by an American or Chinese vendor. The stated reason was control: keeping sensitive government data and systems under French and European jurisdiction, not cost or benchmark scores. That is a coherent trade. A government buying AI for military or classified use has reasons to prize a vendor it can audit, host domestically, and legally compel that a price comparison will never capture. Paying more for a system you actually control is not irrational. It is a different kind of shopping than a business does with its own operating budget.
But a sovereignty bet is not a test of whether Large 4 is open or cheap, and the launch blurs that distinction. As of October 6, the only way to run Large 4 is to pay Mistral per token through a hosted preview on its Mistral Studio platform. The downloadable weights, the thing that actually makes a model open in any usable sense, are not shipping until later this month, with one report putting the date at October 27. Mistral's own announcement and independent coverage from Let's Data Science agree on this: what launched is a closed preview with open weights promised, not an open model.
The performance comparison is thinner than the launch framing suggests, too. Artificial Analysis, the one group that has run Large 4 through its own independent evaluation rather than relying on Mistral's self-reported scores, puts it at 38 on its Intelligence Index, statistically level with GPT-6 Luna and a point behind the open-weight DeepSeek V4.1 Flash at 39. On the Cyber Index, the category Mensch singled out, Large 4 scores 50, tied with the open-weight GLM-5.3-Flash and six points behind the open-weight MiMo-V2.6-Pro at 56. Mensch does have one real data point on his side: on a specific test called CyberGym-E2E-AA, Large 4 leads at 82 percent, ahead of MiMo-V2.6-Pro's 79 percent and GPT-6 Luna's 78 percent. That is one benchmark, not a category win across the board.
The price gap is where the launch framing breaks down completely. Large 4 lists at $1.36 per million input tokens and $4.18 per million output tokens, which works out to $1.13 to clear one task on the Intelligence Index, or $0.57 during Mistral's two-week launch discount. GLM-5.3-Flash clears the same task for about $0.25, and DeepSeek V4.1 Flash for about $0.27, according to Artificial Analysis's own task-cost figures reported alongside the benchmarks above. That is roughly a fifth of Large 4's list price, and still close to half of its discounted rate. A model that is tied on general intelligence, behind the best open rival on cybersecurity, and two to five times the price is not beating the Chinese open-weight field. It is matching pieces of it at a real markup.
None of this makes Large 4 a bad model, and I would not want a reader to walk away thinking so. Landing roughly even with GPT-6 Luna, from a company with nothing like a hyperscaler's balance sheet behind it, is a real achievement, and once the weights actually ship, a business with a genuine data residency requirement will have an option it did not have before. But data residency is a narrow need. Almost none of the companies I talk to about model selection are picking a vendor because of where its headquarters sits. They are picking based on what a task costs to run correctly, and on that question the honest answer today is to wait for the actual open release, then rerun the comparison against whatever GLM and DeepSeek are charging by then, because that field has been repricing itself every few weeks, not every few quarters.
Betting on sovereignty is a government's call to make with government money. A business spending its own still has to shop on price and performance, and right now the Chinese open-weight models are winning that contest. Not because anyone mandated it, but because the price war between them is still the only thing that has reliably pushed the cost of a token down this year.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.