# Mistral Large 3 on the Cotera agent benchmark

> The quiet 5/5. Half the price of GPT-5.5 and didn't whine about JSON once.

Source: https://cotera.co/benchmarks/models/mistral-large-3

---

Mistral AI's flagship dense model out of Paris, the European frontier-model lab's answer to GPT-5 and Claude Sonnet. We tested the November 2026 production endpoint via Mistral's first-party API.

**Score:** 5/5 benchmarks passed · **Cost across the matrix:** $0.267

## Strengths
- JSON discipline. Every run produced a parseable final message conforming to the schema — no preamble, no markdown fences leaking in, no "### Final Output:" prose before the object.
- Tool budget. Used the fewest tool calls of any 5/5 model on three of the five benchmarks. It doesn't double-check itself into the ground.
- Cheap to run. Total spend across the matrix was $0.27 — the cheapest perfect score and roughly 4x cheaper than GPT-5.5, 14x cheaper than Sonnet 4.6, and 17x cheaper than Opus 4.8.
- Predictable shape. Output length and tool-call count varied very little across runs — easy to budget for in production agents.

## Watch-outs
- We tested short-horizon agent jobs (2-9 tool calls per task). We have no data on its behavior across multi-hour, multi-day, or multi-tenant chains where context-management becomes the bottleneck.
- No coding rubric beyond "call the right Stripe primitives." Don't read this as a coding-quality verdict — read it as "can it read docs and produce a working function."
- No adversarial inputs in this suite. We didn't test prompt injection, schema collisions, or partial-tool-failure recovery.