Mistral opened the public preview of Mistral Large 4 on October 6, a 1.05-trillion-parameter mixture-of-experts model (nicknamed "Le Chonk" internally, 49 billion active parameters per token, plus a 1.6B vision encoder) that trained on 3,800 NVIDIA Grace Blackwell GPUs across more than 160 languages. Artificial Analysis dropped it onto Intelligence Index v4.3.2 at a score of 38, which places it 64th of 225 models and well above the class median of 26. Level with GPT-6 Luna (max). One point behind DeepSeek V4.1 Flash (max).

The headline isn't the composite. It's the per-benchmark story underneath.

On Vals' Finance Agent v2, Large 4 posts 54.68% accuracy at $1.20 per test with a 26m 02s latency. GPT-6 Astra, the closed-frontier reference point, lands at 53.54%, a hair lower, but costs $6.82 per test and runs in 11m 33s. GLM 5.3 edges both at 55.84%, and Gemini 4 Argon still rules the leaderboard at 65.40%. The pattern isn't that Mistral won. It's that a mid-priced open-ish model from a European lab now sits inside the error bars of OpenAI's frontier offering on multi-step financial reasoning, at roughly a sixth of the cost.

Legal and cyber tell a similar story. On Harvey's Legal Agent Benchmark, Large 4 passes 15.83% of tasks for $4.18 per test, finishing sixth of 75 models. Kimi K3 trails at 12.92%, GLM-5.3 at 8.33%, GPT-6 Astra at 5.42%, Claude Opus 5.5 at 3.75%. The absolute pass rates are low; the ordering is what matters.

The composite Vals Index v2.1 keeps everyone honest: Large 4 scores 48.05% ±1.11 and ranks 32nd of 44, with GPT-6 Astra at 63.13%. General-purpose breadth still belongs to the frontier. Verticals are up for grabs.

For small-business owners deciding what tier of model infrastructure is "good enough" to power personalized prospect research and outreach drafting, this is the actionable read. Capable multi-step business reasoning no longer requires frontier-priced inference. LemonLime, the AI that gets you customers, prepares exactly that kind of work, lead qualification, per-prospect outreach, missed-opportunity surfacing, proactively, every morning. Each compression cycle in model pricing deepens what a done-for-you service can hand to an owner before 9am.

The cost-performance curve has been bending for a while. GPT-6.1 Sol arrived at one-fifth of Astra's price while outscoring Opus 5.5 earlier this month. Z.ai's 320B-A18B MIT-licensed GLM-5.3 Flash previewed as a mystery model in August. Large 4 is the next tick.

One asterisk. The open weights, promised as Mistral-Large-4.0-1T05-A52B, aren't out. OrcaRouter checked Mistral's Hugging Face organization on October 6 and found nothing newer than July. The planned drop is October 31. Until then, "open" is a roadmap, not a repository.

Mistral Large 4 Preview Lands at 38 on Intelligence Index, Edges GPT-6 Astra on Finance Agent v2

Sources