Mistral shipped Large 3 in February 2026 with full EU data residency, native function calling, and pricing aimed at undercutting frontier US labs. For European enterprises and any organization where US data flow is a problem, it's the most pragmatic frontier option in 2026 — but it's not without tradeoffs.
What changed in 2026
- EU residency is now real and verifiable. Models hosted in France, Germany, and Switzerland with audit trails. Mistral can sign DPAs that explicitly exclude US data flow.
- Function calling matured. Parallel calls, structured outputs, and JSON mode now match OpenAI's reliability on most workloads.
- Pricing dropped 30%. $2/M input · $6/M output for Large 3, down from Large 2's $3/M · $9/M.
Quality vs frontier
On English-only benchmarks, Mistral Large 3 trails GPT-5 and Claude Opus 4.7 by roughly 4-6 points across MMLU, HumanEval, and reasoning suites. That gap is real but not always decisive — for many production workloads, Large 3 is within the "good enough" band, especially when paired with strong prompt engineering and RAG.
On European-language workloads, it's a different story. French, German, Spanish, and Italian benchmarks favor Mistral by 3-5 points over OpenAI and Anthropic. If your customers are European, this is decisive.
Function calling in production
We migrated a customer-support agent from GPT-4o to Mistral Large 3 in March. Tool-call accuracy at 95% (vs 97% on GPT-4o), and JSON-mode validity at 99.2% (vs 99.6%). Within tolerance for most production agents. Latency improved by ~30% for European users due to regional hosting.
Pricing and deployment
| Model |
Input ($/M) |
Output ($/M) |
Context |
| Mistral Large 3 |
$2.00 |
$6.00 |
256K |
| Mistral Medium 3 |
$0.40 |
$2.00 |
256K |
| Mistral Small 3 |
$0.10 |
$0.30 |
128K |
| Le Chat Enterprise |
flat (€20-50/seat) |
— |
bundled |
Le Chat Enterprise is the productized chat product (think ChatGPT Enterprise) — increasingly winning EU enterprise deals where data residency forecloses ChatGPT. Pricing is per seat with unlimited usage.
Where Mistral wins
EU-regulated data. GDPR, DORA, NIS2 audits go smoother when data never leaves the EU. Mistral signs DPAs that explicitly exclude US persons.
Edge / on-device. Mistral Small 3 (8B) runs well on consumer hardware. The Q4 quant is ~5GB and runs on M-series Macs at 30+ tok/s.
Government and defense. Mistral has the cleanest sovereign story in Europe, with deals with the French and German governments active in 2026.
Where Mistral loses
Bleeding-edge reasoning. GPT-5 thinking and Opus 4.7 thinking still beat Mistral on the hardest reasoning benchmarks.
Multimodal. Vision is okay; voice is not yet competitive with GPT-5's native voice.
Long context. 256K is workable but trails Gemini 3 Pro (2M), Llama 4 Scout (10M), and Opus 4.7 (1M).
When to choose Mistral
Choose Mistral when EU residency is a real requirement, your workload is multilingual European, your workload is heavy on Le Chat Enterprise as a productivity tool, or you want a frontier alternative outside the US duopoly. Skip Mistral if your workload demands the absolute highest reasoning ceiling.
FAQ
Can I run Mistral Large 3 on Azure?
Yes. Mistral Large 3 is on Azure AI Foundry with EU regions and full enterprise compliance.
Are the open-weights models still being released?
Yes — Mistral Small 3 weights are open under Apache 2.0. Mistral Large 3 remains closed-weights and API-only.
How does it compare on coding?
Codestral 3 (the coding-specialized variant) is competitive with GPT-5 and Opus on Python, JavaScript, and TypeScript. Slightly behind on rarer languages.
Where to go next
For related deep dives see GPT-5 vs Claude Opus 4.7, DeepSeek R2 review, and AI regulation 2026 — the EU AI Act.