Scale · founder · 6 min read

Mistral Large 4: a cheap second-place coding model, with a catch

Mistral's Large 4 preview ranks second in a blind coding test at a fraction of Claude Opus 5's price. What it means for your tools, and what to wait for.

Published October 8, 2026 ·

mistral mistral-large-4 models pricing open-weights founders

On October 6, Mistral put out a preview of Large 4, its biggest model yet: roughly a trillion parameters in total, with 49 billion active on any given request, and natively multimodal. It’s available now through Mistral Studio’s API. Open weights are promised by the end of October.

Most coverage will be about the parameter count. That’s the least useful number in the announcement. Here’s what matters if you’re building with AI tools rather than benchmarking them.

What Mistral actually claims

The coding numbers, all reported by Mistral and in two cases run through Artificial Analysis’s coding index:

• DeepSWE v1.1: 61.7%
• SWE-Atlas-QnA: 59.4%
• Terminal-Bench 4: 28.3%
• Combined Coding Agent Index: 49.8

For context, Grok 4.6 scored 65.9% on DeepSWE v1.1 in August, and GPT-5.6 Sol Max sat at 73%. So on that benchmark Large 4 is competitive, not leading.

The more interesting result is a blind human evaluation run with Surge AI on a 1-to-5 scale. Large 4 scored 3.74, second of five models, behind Claude Opus 5 at 4.22. Blind human ratings are harder to game than benchmark suites, and “second place, clearly behind first” is a believable result for a model like this.

One caution: these are Mistral’s own published figures for a preview. Nobody independent has had weeks with it yet.

The price is the story

According to Mistral’s announcement page, preview pricing is $1.36 per million input tokens and $4.18 per million output tokens. That’s under half what Claude Opus 5 costs, and cheaper than Grok 4.6’s $2/$6 base rate. Another outlet covering the launch didn’t list pricing at all, so treat the exact figures as Mistral’s, and check the Studio pricing page before you budget around them.

If those numbers hold, Large 4 lands in a useful spot for anyone calling models directly: close enough to the frontier that most day-to-day tasks work, cheap enough that a long agent run doesn’t make you flinch. That matters most if you’re building a product that calls a model on every user action.

What this changes for you today

For most readers, honestly, very little. Large 4 is an API model in preview. It isn’t a tool you open and use.

• If you use Lovable, Bolt, Replit or v0: nothing changes. These tools choose their own models, and none has announced Large 4 support. Mistral’s page doesn’t say whether it’s coming to Mistral Vibe either.
• If you use Cursor or Claude Code: nothing yet. Watch the model picker over the next few weeks.
• If you’re building a product on a model API: this is worth a test, especially for cost-sensitive features. Run your own prompts through it. Benchmarks measure other people’s tasks.

Why the open weights matter more than the preview

The part to watch is the end-of-October release of weights, plus the architecture details and training methodology Mistral says will come with them. Open weights mean other companies can host the model, which tends to push prices down and lets tools offer it as a cheap option. It’s also how a model ends up inside the tools you already use.

Two things aren’t clear yet. Mistral hasn’t said what license the weights will ship under, and that decides whether you can use the model commercially without friction. And a trillion-parameter model isn’t something you run on a laptop, so “open” here means open to hosting providers and well-funded teams, not to you.

The European angle

The preview runs on Mistral’s own European datacenters, and the company says it will offer a European deployment it operates end to end. If you sell to customers who ask where their data is processed, that’s a real selling point, and it’s the same wedge Mistral used when it rebranded Le Chat as Vibe. It’s less relevant if you’re a US startup with no such requirement.

What to do this week

Nothing urgent. A reasonable plan:

1. Don’t switch anything on a preview. Previews change pricing, limits and behavior without much notice.
2. If you call a model API in your product, set up a small test with ten of your real prompts and compare Large 4 against what you use now on quality and cost.
3. Check back at the end of October for the license and the list of hosting providers. That’s when this becomes either a genuine cheap option or just another leaderboard entry.

Bottom line

Large 4 is a credible second-tier frontier model at a lower price, which is exactly what keeps pressure on model pricing across the category. It won’t change which app builder you pick. It may change what your own product costs to run, once the license and the hosting options are known.