Mistral Large 4 is worth a test this month if you do document-heavy legal work, need EU-hosted inference, or want image input at a low sticker price. It is not yet the open model the launch suggests: as of 7 Oct 2026 there are no weights and no license, the 50% preview discount has no end date, and its long answers push the per-task cost at list price to about Claude Sonnet 5.5’s level.
Mistral released Large 4 on 6 October 2026 as a public preview, nicknamed “Le Chonk”. For now it is API-only through Mistral Studio. Mistral’s model page describes a mixture-of-experts model with 1.05 trillion total parameters, 52 billion active per token and a 1.6B vision encoder. It reads text and images and answers in text. Some early coverage printed 49 billion active parameters and 4,000 GPUs; Mistral’s own blog says 52 billion and 3,800 Grace Blackwell GPUs, so this review uses Mistral’s figures.
The launch sells the model on three ideas: it’s cheap, it’s at the frontier, and it’s open. Each one is partly true. Here is which part.
“Cheap” is true on the price card, less so on the bill
Mistral’s API pricing page (checked 7 Oct 2026) shows Large 4 at a sale price that is exactly half its list price on every line. No end date is printed next to it.
| Model (Mistral API) | Input / 1M tokens | Cached input / 1M | Output / 1M tokens |
|---|---|---|---|
| Large 4, list price | $1.36 | $0.14 | $4.18 |
| Large 4, current sale | $0.68 | $0.07 | $2.09 |
| Large 3 | $0.50 | $0.05 | $1.50 |
| Z.ai GLM 5.3 (hosted by Mistral) | $1.40 | $0.14 | $4.40 |
Two things in that table are easy to miss. Large 4’s sale price is still about 1.4 times Large 3’s, on input and output alike. And Mistral sells Z.ai’s GLM 5.3 on the same page at almost exactly Large 4’s list price. If the discount ends quietly, Mistral’s new flagship and a Chinese open model will cost you nearly the same per token.
The sticker doesn’t decide what you pay. Token count does. Artificial Analysis ran the preview through its Intelligence Index v4.3.2 and recorded 200 million output tokens, about two and a half times the 81 million median for its comparison group. It calls the model “very verbose”. At list prices that works out to $1.13 per index task.
| Model (Artificial Analysis setting) | Intelligence Index v4.3.2 | Cost per index task | Output tokens on the index |
|---|---|---|---|
| Large 4 Preview (reasoning) | 38 | $1.13 | 200M |
| GLM-5.3 (max) | 45 | $2.01 | 210M |
| Claude Sonnet 5.5 (high) | 47 | $1.12 | 52M |
| GPT-6.1 Sol (high) | 50 | $0.32 | 25M |
| Large 3 (non-reasoning) | 9 | $0.03 | 9.8M |
Read across the rows and the picture changes. At list price, a Large 4 task costs the same as a Claude Sonnet 5.5 task at high effort, for a score nine points lower. GPT-6.1 Sol at high effort scores 12 points higher for under a third of the cost. If the sale holds for every token type, as the price card shows, the same token use would come to roughly $0.57 per task. That’s our arithmetic, not a figure Artificial Analysis published, and it’s still well above Sol.
Reasoning models usually bill their thinking as output tokens, whether you read that thinking or not, which is why our guide to reasoning-token pricing matters more here than the headline rate. The Sonnet and Sol rows come from the same index version; we went through them in detail in Sonnet 5.5 vs GPT-6.1 Sol. One composite is still one view, though. On legal agent work, covered next, the order flips hard.
“Frontier” depends on which leaderboard you open
Mistral published a long list of its own results. Independent trackers had numbers up within a day. Where the two overlap:
| Test | What Mistral says | Independent result (as of 7 Oct 2026) |
|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 28.3% | 22.73% on Vals AI, 21st of 45 |
| Vals Finance Agent v2 | Beats GPT-6 Astra | True by 1.1 points (54.68% vs 53.54%), but 22nd of 76. GLM 5.3 scores 55.84%; GLM 5.3 Flash scores 57.85% at $0.05 a test |
| Harvey’s Legal Agent Benchmark | Best among open models | 15.83%, 6th of 76 (tied with Grok 4.6), ahead of every Claude and OpenAI model on the board |
| Overall standing | Best open-weight model from the US or Europe | Artificial Analysis: 38, #64 of 225. Vals Index: 48.05%, 33rd of 45 |
The Terminal-Bench gap is more than five points. Different harnesses and settings can explain some of that, and Mistral’s post doesn’t say which setup produced its figure.
The finance claim is a true sentence doing misleading work. Large 4 does edge out GPT-6 Astra on Vals’ Finance Agent leaderboard, but Astra sits 29th on that particular test. Twenty-one models score higher than Large 4, including GLM 5.3 Flash at roughly a twenty-fourth of the cost per test. Beating the one rival you picked isn’t the same as leading.
The legal result is the real one. On Harvey’s agent benchmark, a task only counts if every grading criterion passes. Large 4 resolved 15.83% of tasks. Claude Opus 5.5 managed 3.75%, and GPT-6 Astra and GPT-6.1 Sol 5.42% each, with Opus costing about five times as much per test ($21.38 vs $4.18). Keep the scale in mind: even the leader, Muse Spark 1.2, fully resolves about a quarter of tasks. Large 4 gets roughly one in six completely right, so treat its output as a draft for a lawyer, not a finished document.
Benchmark scores aren’t legal or financial advice; have a qualified person check anything you’ll rely on.
Ranks move by the day, too. Vals’ model page for Large 4 showed it 32nd of 44 on the Vals Index early on 7 October and 33rd of 45 later the same day, after Claude Haiku 5.5 appeared on the site. We’ve covered why the same model lands in different places across trackers in our AI model rankings explainer.
“Open” is a promise without the paperwork
Mistral’s blog puts it plainly: “Weights drop end of this month.” A 27 October date has appeared in press coverage, but it isn’t on Mistral’s page, so treat it as reported. For a business user the bigger gap is the license. None has been published. Large 3 shipped under Apache 2.0, which allows commercial use with few strings, and nothing guarantees Large 4 gets the same terms. Mistral’s documentation and Vals already file the model as open, while Artificial Analysis lists it as proprietary because nothing can be downloaded yet. Same model, two labels, same week.
Until the weights ship, Mistral says it is red-teaming the model with security firms, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber abilities. That’s a sensible step. It also means the model you test today may not match the weights released later, because Mistral says its reinforcement-learning run is still going and the model keeps improving.
And open won’t mean “runs in your office”. At 1.05 trillion parameters, the weights alone come to roughly 525 GB at 4-bit precision (our arithmetic: 1.05 trillion parameters at half a byte each), before any memory for long contexts. That’s a multi-GPU server. We hit the same wall with Xiaomi’s MiMo V2.6 Pro, and our local AI models guide shows what laptop-class hardware can realistically handle. For a small team, open weights pay off in other ways: other hosts can compete on price, and you can move providers later without rebuilding around a different model.
Where Mistral Large 4 earns a test slot
Set the labels aside and some real strengths remain. Independent results and Mistral’s own claims are marked separately here.
- Legal and document-heavy agent work (independent). The Harvey result above comes from Vals, not Mistral, and it’s well ahead of the US closed models on that test.
- Quick start, long finish (independent). Artificial Analysis measured 116 output tokens per second and 1.46 seconds to the first token on Mistral’s API. Long reasoning still stretches agent jobs: Vals logged about 26 minutes per Finance Agent test and 39 minutes per Harvey task.
- Images in, not just text (independent). GLM-5.3, the higher-scoring open model in the cost table above, only reads text.
- Security work (Mistral’s claim). Mistral says Large 4 ranks in the top five on Artificial Analysis’ Cyber Index and scores 82% on a vulnerability-reproduction test where some closed models score near zero because they refuse. It also says the model refuses malicious cyber prompts more often than other open models. We read all of this only on Mistral’s page.
- Where the data sits (Mistral’s claim). Mistral says the preview runs in its own European datacenters, with a European deployment it operates end-to-end under EU law. If data residency is a client requirement, that alone may put it on your shortlist. Check the contract terms, not the blog post.
Coming from Large 3? Don’t just swap the model ID
Large 3 is a non-reasoning model. On Artificial Analysis it scores 9, uses 9.8 million output tokens across the whole index and costs about $0.03 per task. Large 4 scores 38 and uses roughly 20 times as many output tokens. For classification, short rewrites or pulling fields out of invoices, that quality jump may buy you nothing while the per-task cost climbs about 38-fold at list price (based on Artificial Analysis’ rounded figures).
Press reports describe two reasoning settings, “none” and “high”. Artificial Analysis has only published the reasoning version so far, so nobody independent has priced the no-reasoning mode. Run your real jobs both ways before switching. Our method for comparing models on your own work takes about an afternoon and counts cost per usable answer, which is the number that matters with a model this talkative.
One loose end: Artificial Analysis now marks Large 3 as deprecated on its own site, but Mistral’s pricing page still lists it with no retirement date shown.
Who should do what before the weights land
Test it now if you run document-heavy legal or contract work, if EU hosting is something your clients ask for, or if you need image input from a non-US provider. Budget at the list price, not the sale price, so a quiet end to the discount doesn’t break your numbers.
Wait for the license if your plan depends on self-hosting, fine-tuning or reselling access. Until the terms are public, you don’t know what you’re allowed to do with the weights. Our DeepSeek V4 vs GLM-5.3 comparison shows how “open” licenses can carry conditions you only find by reading the text.
Look elsewhere first if coding agents are your main use, or you want the most general-purpose score per dollar. On Artificial Analysis’ index, GPT-6.1 Sol scores higher for under a third of the per-task cost, and Large 4’s coding figure shrank on Vals’ Terminal-Bench run. GLM-5.3 also scores higher there but costs more per task; our GLM-5.3 review covers that model and its unusual license.
What we couldn’t confirm
- When the 50% preview discount ends. Mistral’s price page shows no date.
- The license terms, and the exact weights date. Mistral says end of October; 27 October is press-reported.
- The context window. Mistral’s model page says 1M tokens; Artificial Analysis lists 524K and Vals 512K. Test with your longest real document before relying on the larger figure.
- Active parameters and training hardware. Mistral’s pages say 52B active and 3,800 GPUs; early coverage said 49B and 4,000.
- How the “none” reasoning setting performs or what it costs per task. No independent numbers yet, and the two-setting description comes from press reports.
- Mistral’s cybersecurity figures (top-five Cyber Index ranking, 82% on the vulnerability test, 93% on Cybench). We read them only on Mistral’s page.
- Whether Large 4 is in Mistral’s consumer chat app. The launch post points only to Studio and the API.
- Whether the released weights will score the same as today’s preview, since Mistral says training is still running.
FAQ
Is Mistral Large 4 open source?
Not yet. As of 7 October 2026 it is an API-only preview. Mistral says weights will be released by the end of October, but no license has been published, so you can’t yet know what commercial use will be allowed.
How much does Mistral Large 4 cost?
List price is $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. Mistral currently charges half of each, with no stated end date. Because the model writes a lot of reasoning tokens, Artificial Analysis measured $1.13 per task on its index at list prices.
Can I run Large 4 on my own computer?
No. Even after the weights are out, 1.05 trillion parameters need roughly 525 GB for the weights alone at 4-bit precision, which means a multi-GPU server. Most small teams will use it through Mistral’s API or another host.
Is Large 4 better than GLM-5.3?
It depends on the test. GLM-5.3 scores higher on Artificial Analysis’ index (45 vs 38) but costs more per task there ($2.01 vs $1.13). On Harvey’s legal agent benchmark, Large 4 is well ahead (15.83% vs 8.33%). Large 4 also accepts images; GLM-5.3 is text-only.
Is Large 4 good for coding?
It’s competent, not a leader. Mistral reports 28.3% on Terminal-Bench 4.0; Vals measured 22.73%, 21st of 45. If coding agents are your main use, test it against your current model before switching.
Should I move from Mistral Large 3?
Only for tasks that need reasoning. Large 3 is still listed at $0.50/$1.50 per million tokens and costs far less per task for simple jobs. Run a side-by-side on your own workload first.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
