Mistral announced Mistral Large 4 on October 6, 2026, and the number I stopped on was $1.36. That is the published API price per million input tokens, with output at $4.18, for a model Mistral describes as 1 trillion parameters with 49 billion active. I route long client drafts through Claude every day and keep sensitive work on a local Ollama stack, so a large open-weight model with a European hosting option lands right on the line I care about. I have not used Mistral Large 4 myself, so this is a read from the sidelines, built on Mistral’s own announcement.
In this article
- What Mistral actually shipped on October 6
- Why Mistral Large 4 is a sovereignty story before a benchmark story
- How to read the launch benchmarks
- Where it could sit next to Claude and a local stack
- Five checks before routing any client work to it
Mistral Shipped a Preview Today and Promised Weights This Month
The short version: Mistral Large 4 is live as a public preview, and the open weights are a promise with a date, not a download yet. Mistral says the preview is available through Mistral Studio, and states plainly: “We will release the weights by the end of the month.”
The launch facts that frame everything else:
- 1 trillion total parameters, 49 billion active, in a mixture-of-experts design
- A hybrid instruct-and-reasoning model that is natively multimodal
- Support for more than 160 languages, including all official EU languages
- API pricing of $1.36 per million input tokens and $4.18 per million output tokens
- Training on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters
Two things are missing from the announcement, and they matter. It does not state a license for the coming weights, and it does not give a context window figure I could confirm. Until both are published, any plan that depends on running Mistral Large 4 yourself is a plan built on a gap.
Mistral Large 4 Is a Sovereignty Story Before a Benchmark Story
The benchmarks get the charts, but the hosting sentence is the real news for client work. Mistral says the model “will be available across multiple regions worldwide, including a European deployment that Mistral operates end-to-end, independently of other digital service providers and under European law.”
That is a specific claim, and it answers a question I have been asking for months. When I wrote about DeepSeek V4 Flash pricing, my position was that cheap frontier-class models deserve a fair look, but client data does not get routed to them while data sovereignty is unclear. Mistral Large 4 is the first launch this autumn where the sovereignty answer is the pitch itself.
There is a difference between a pitch and a contract, though. “Under European law” tells you which rules apply. It does not tell you the retention period, whether prompts can feed training, or what the data processing terms say. Those live in documents a solo consultant should read before the first paid call, not after.
The weights add a second path. If Mistral Large 4 ships with a permissive license, the most cautious client could, in principle, run it on infrastructure they control. In practice, 1 trillion parameters is not a laptop model, and nobody should plan around a local build of this size.
The Launch Benchmarks Are Strong but Self-Reported
Mistral Large 4 posts solid numbers, and every one of them comes from Mistral. The announcement compares it against Claude Opus 5, GPT-6 Astra and DeepSeek V4 Pro, and claims it is “significantly outperforming any open-weight model developed in the US or Europe.”
The figures worth noting for a solo stack:
- DeepSWE v1.1 at 61.7% and SWE-Atlas-QnA at 59.4% for agentic coding
- Terminal Bench 4.0 at 28.3%, a reminder that terminal-heavy agent work is still hard
- AutomationBench at 59.9% for agentic workflows
- Cybench at 93% and B3 attack resistance at 93.3% on the security side
A cheaper model with a clear legal home is worth more to client work than a slightly smarter model with a vague one.
What I would not do is treat these as a ranking. Launch-day charts are chosen by the company launching, and independent runs usually arrive weeks later. The useful signal is narrower: Mistral is positioning Mistral Large 4 as an agentic and coding model, not only a chat model, and that is the claim worth testing first.
Mistral also notes that reinforcement learning ran at “33 billion tokens per day” during training. It is an interesting engineering detail, but it says nothing about how the model handles a messy client brief.
Mistral Large 4 Could Fill the Gap Between Claude and Ollama
Where it fits is a middle lane, and that lane is real in my setup. Claude handles long drafts and strategy documents. Ollama with Hermes 4 handles anything containing client names, numbers or personal data, because it never leaves my machine. The weak spot is work that is too heavy for a local model but too sensitive for a default cloud setting.
That is the slot Mistral Large 4 is aiming at. A large model, priced below the flagships, with an EU deployment the vendor operates itself, is exactly what a consultant with European clients might want for that middle lane. I laid out my rules for client data on local AI earlier, and a hosted European model would need to clear the same bar from the other side.
It will not replace the local stack, and size is the reason. My notes on GGUF quants cover what a single machine can realistically run, and a model with 49 billion active parameters out of 1 trillion is far beyond a solo desk. Even after the weights land, the realistic options are Mistral’s own hosting or a third-party host.
Five Checks Before Routing Client Work to Mistral Large 4
My conclusion is that Mistral Large 4 is worth a test on public material, not a workflow change yet. If I tried it this month, I would run these checks in order:
- Read the data terms first. Find the retention and training-use language for the EU deployment before sending a single prompt.
- Wait for the license. The weights are due by the end of October; the license decides whether self-hosting is even an option.
- Price a real job. Take one recent long draft, count its input and output tokens, and price it at $1.36 and $4.18 per million.
- Test in the language you sell in. The 160-language claim matters most if your clients write in something other than English.
- Run one agentic task. The benchmarks lean on coding and automation, so give it a small file-cleanup job and compare it with Claude Code.
The preview is the easy first step. Mistral’s announcement page points to Mistral Studio, so you can see the output style without committing anything sensitive. If the data terms hold up, a single non-confidential client task is the next experiment.
For me, Mistral Large 4 is the most useful kind of launch: a clear price, a clear legal home, and a weights date on the calendar. It will not replace Claude for long-form strategy, and it will not replace Ollama for data that must stay on my machine. But a European middle lane could close a gap I have worked around for months. I will read the data terms, wait for the license, and report back once the weights ship.
Sources
AI-assisted research and drafting. Reviewed and published by ToolMint.