Mistral Large 4 vs Opus 5.5: index 26 vs 58, $2 vs $10 a task

Mistral Large 4 vs Opus 5.5 is not a close contest on intelligence, but it is a lopsided one on price. Artificial Analysis scores Claude Opus 5.5 at 58 on its Intelligence Index and Mistral Large 4 at 26. Mistral's API charges $0.70 per million input tokens against $4 for Opus 5.5. Mistral released the open weights on October 29 under the nickname "Le Chonk", and the model is the best-placed open option in several of its own tests. In independent ones, it sits well behind Anthropic's flagship.

This comparison is based on the companies' docs and published results, not our own testing. Prices are as of October 29, 2026.

Mistral Large 4 vs Opus 5.5 at a glance

Mistral Large 4 is a mixture-of-experts model, which runs only part of its network for each token. It takes images as input and has low, medium and high reasoning modes, with high as the default. Opus 5.5 came out on September 22 and is available as claude-opus-5-5 on the Claude Platform, AWS, Google Cloud and Azure.

Independent benchmarks: Opus 5.5 leads by a wide margin

Artificial Analysis, an independent testing firm, lists Mistral Large 4 (High) at 26 on its Intelligence Index. That is above the median of 23 for open-weight models it compares against. It lists Opus 5.5 at 58, second of 54 models in that listing.

The index is a blend of tests, so the gap is not one task. It does say Opus 5.5 is the stronger model for hard, open-ended work. The cost line tells the other half: running the index costs about $2.06 per task on Mistral and $10.04 on Opus, roughly five times more. Opus costs more per token, but Mistral is not nine times cheaper per finished job.

What Mistral claims

Mistral's own announcement reports results mostly on agent and coding work. It says Large 4 scores 41% on Terminal-Bench 4.0, 19.7% on Cognition's private FrontierCode 2.0, and 48% on the Artificial Analysis Coding Agent Index. It calls the first two the best among open-weight models. In a blind human-preference test from Surge, Mistral says raters preferred it to GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and that only Claude Opus 5.5 ranked higher among the models evaluated.

Anthropic reports 66.4% for Opus 5.5 on Terminal-Bench 4.0, at max effort. That is 25 points above Mistral's number, though the two companies may not use the same test setup. Anthropic's FrontierCode figure (54.4%, v1.1 Main) is a different version from Mistral's, so the two should not be compared. All of these are company-reported.

Price: where Mistral wins

On Anthropic's pricing page, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Mistral's list price is about 5.7 times lower for input and 9.5 times lower for output. Our earlier article covers what Opus 5.5 costs and what changes the bill.

The weights are the bigger difference. Mistral's license allows use, modification, fine-tuning and commercial use, with one exception: a company that sells the model as a hosted service needs a separate license from Mistral. Mistral names vLLM, SGLang and TensorRT-LLM as ways to serve it. It does not state hardware requirements, and a model of this size needs a serious GPU cluster. Self-hosting only beats the API at steady, high volume.

Which one to pick

Pick Opus 5.5 when answer quality on hard agentic work matters more than the bill. Pick Mistral Large 4 when you need to run the model yourself, keep data in your own infrastructure, fine-tune it, or when the price of Opus rules it out. A cheap model that needs a second pass can cost more than it saves, so test on your own tasks first.

Mistral's announcement is covered in more detail in our news article on the open-weights release. Artificial Analysis may update its scores for the final release, so recheck its pages before you decide.