🔍 Read the full analysis: What Mistral Large 4’S AI Frontier Gap Says About The Race on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get the latest gadgets delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Mistral launched Large 4 in public API preview on October 6, but the model scored 38 on Artificial Analysis’s Intelligence Index, below several leading U.S. and Chinese systems in the October 7 comparison. The result informs a reviewer’s judgment that developers should look to stronger alternatives for demanding agentic work, although the benchmark is not a direct test of every real-world workload.

Large 4 is Mistral’s largest model to date: a mixture-of-experts system with one trillion total parameters and 49 billion active parameters. It accepts text and images. Mistral says it trained the model on its own infrastructure in Europe and is continuing to improve it. The model is currently available through a preview API; its weights are scheduled for release later in October and are not publicly downloadable as of the October 7 assessment.

Artificial Analysis’s Intelligence Index assigns Large 4 Preview a score of 38. In the comparison provided by the review, Anthropic’s Claude Opus 5.5 scored 58, Google’s Gemini 4 Argon 53 and OpenAI’s GPT-6.1 Sol 52. Chinese systems Z.ai GLM-5.3 and Moonshot AI’s Kimi K3 scored 45 and 44, respectively. DeepSeek V4.1 Flash scored 39. OpenAI’s GPT-6 Luna also scored 38, while Canada’s Cohere Command A+ scored 13.

The review describes those figures as a dated benchmark snapshot, not a forecast of task success. Models were evaluated at different named reasoning settings, so the comparison was not conducted under identical compute budgets. The review also reports roughly 512,000 tokens of context capacity for Large 4; that measures how much input can fit, not whether the model reasons correctly over it.

The Gap for Complex AI Work

The score matters because developers choosing models for long, agentic tasks need more than a large parameter count or context window. Agentic systems plan, call tools, interpret results and carry decisions through multiple steps. An unsupported assumption early in that chain can affect later actions, while a fluent final answer may not reveal where the process went wrong.

The review’s author says they would not select the current preview for demanding autonomous work when higher-scoring alternatives are available. They report encountering hallucinations in their own use and say that this experience reduced their confidence. That is personal observation, not a controlled comparison, and it does not establish how often Large 4 hallucinates relative to other models. The practical issue is whether it can complete a particular workload accurately with acceptable supervision.

The results also show why “frontier” is not a single, settled category. Mistral’s score matches GPT-6 Luna at maximum reasoning effort and sits close to DeepSeek V4.1 Flash, but it remains below other listed U.S. and Chinese systems. The index offers one aggregate view; developers still need to test their own coding, research or professional tasks before choosing a model.

A Preview, Not the Weight Release

The release is a step in Mistral’s effort to build European AI capacity. The company says Large 4 was trained on its own European infrastructure, and its trillion-parameter total makes it the largest model Mistral has introduced. Those facts describe the model and its development; they do not establish that it matches the strongest available systems on performance.

The distinction between today’s API preview and the planned weight release matters to developers. On October 7, the preview could be accessed through Mistral’s API, while downloadable weights had not yet been released. Mistral says the model is still being improved, so results from this version may not describe later releases. The reviewer’s assessment is explicitly about the current preview, not a final judgment on future versions.

Artificial Analysis’s comparison also needs careful reading. Its listed scores use different reasoning settings, and the index does not measure every capability or the reliability of a model on a given company’s workflow. The review says DeepSeek V4.1 Flash has approximately comparable benchmark intelligence at a much lower measured cost per task, but the provided material does not give the underlying cost figures. Cohere Command A+ scored below Mistral, showing that the comparison does not place every competitor ahead.

“I would not choose it for demanding agentic work or long tasks when stronger models are available.”

— Thorsten Meyer, the review’s author

Performance Beyond the Index

It remains unclear how Large 4 performs across specific workloads, including long coding and research tasks, because the cited Intelligence Index is an aggregate measure rather than a direct evaluation of those tasks. The review’s account of hallucinations reflects one user’s experience; no controlled comparative hallucination results are provided. The source material also does not include enough detail to verify the reported cost comparison with DeepSeek V4.1 Flash.

Future performance may change as Mistral continues to improve the preview and releases the model weights. It is not yet clear when in October the weights will become available, whether they will match the API version, or how later versions will score. The benchmark snapshot may also change as other models and evaluation results are updated.

The Weight Release and Retesting

Mistral has scheduled the Large 4 weights for release later in October. Developers will be able to assess that release once it is available, subject to the terms and technical details Mistral provides. Because the current assessment covers an API preview, results from the released weights should not be assumed to be identical without testing.

The next useful evidence will include updated benchmark results and workload-specific evaluations of accuracy, tool use, constraint-following and the amount of human supervision required. Developers comparing providers should also check pricing and test models under comparable settings. Until those details are available, the current assessment is a time-stamped view of a preview—not a final verdict on Mistral’s future models.

Key Questions

What is Mistral Large 4?

It is Mistral’s largest model to date, a one-trillion-parameter mixture-of-experts system with 49 billion active parameters. It accepts text and images and was launched as an API preview on October 6, 2026.

How did Large 4 score against the models listed?

Artificial Analysis gave Large 4 Preview an Intelligence Index score of 38. That is below several listed U.S. and Chinese models, equal to GPT-6 Luna in the comparison, and above Cohere Command A+. The comparison uses different reasoning settings and is a dated snapshot.

Does a score of 38 mean Large 4 will fail at agentic tasks?

No. The score is an aggregate benchmark result, not proof of failure on a particular coding, research or tool-using task. Developers need workload-specific testing to judge accuracy and reliability for their use.

Are Mistral Large 4’s weights available?

Not as of the October 7 assessment. Mistral has made the model available through a preview API and has scheduled a weight release later in October, but a specific release date was not given.

What remains unverified about the criticism?

The review’s hallucination concerns come from the author’s personal use, not a controlled comparison. The cited material also does not provide detailed cost figures or results across specific long-task workflows, so those claims require further testing.

Source: ThorstenMeyerAI.com