🔍 Read the full analysis: How Far Behind The AI Frontier Is Mistral Large 4? 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 in a comparison dated October 7—below several leading U.S. and Chinese models. That gap raises questions about its suitability for demanding agentic work, though the result is a benchmark snapshot of a preview, not a verdict on the model’s eventual performance.

Mistral describes Large 4 as its largest model to date, a mixture-of-experts system with one trillion total parameters and 49 billion active parameters. The preview accepts text and images through an API. The company said it trained the model on its own infrastructure in Europe and is continuing to improve it. Mistral has scheduled a release of the weights for later in October; at the time of the source report, they were not available to download.

Artificial Analysis’s October 7 Intelligence Index comparison gives Claude Opus 5.5 a score of 58, Gemini 4 Argon 53 and GPT-6.1 Sol 52. Chinese models GLM-5.3 and Kimi K3 scored 45 and 44, while DeepSeek V4.1 Flash scored 39. Mistral Large 4 Preview’s 38 matched GPT-6 Luna at maximum reasoning effort and was above Cohere Command A+’s 13. The listed models were evaluated at different reasoning settings, so this is not a comparison under identical compute budgets.

The source author, Thorsten Meyer, says his own use of the preview included hallucinations and that he would not select it for demanding, long-running agentic tasks when stronger alternatives are available. He presents that as personal experience and a current judgment, not the result of a controlled reliability study. Artificial Analysis’s aggregate score also does not directly measure success on every coding, research or professional workflow.

Benchmark Gaps Shape Model Choice

For developers choosing a model for multi-step work, benchmark standing can help narrow the field, but it cannot settle the decision. An agentic system may plan, use tools and carry findings through several steps; an unsupported assumption early on can affect the final result. The source author’s concern is that lower aggregate benchmark performance, alongside his reported experience, gives him less reason to trust this preview with extended tasks without close supervision.

The scores indicate a specific competitive gap, not a universal inability. Mistral is 10 index points below GLM-5.3 and 20 below Claude Opus 5.5 in this snapshot. Index points are not percentages of intelligence or predictions of task success. Mistral also scores above Cohere Command A+ in this comparison, so the evidence does not support saying that every competing lab is ahead.

For European AI capacity, the launch still has relevance: Mistral says the model was trained on its own European infrastructure. But that development and the question of whether it is the best model for a particular workload are separate. Developers will need to weigh performance, cost, reliability, deployment options and their own test results.

A Preview, Not the Weight Release

The timing matters: Large 4 is currently a preview API, while the announced weight release is scheduled for later in October. Conclusions drawn now apply to the preview version and may need revisiting if Mistral makes changes before or after releasing the weights. The source material does not provide a date for a final release or a complete account of what will change.

Artificial Analysis reports a context capacity of roughly 512,000 tokens. That indicates how much input the model can accept, not whether it can reason accurately over all of it. Similarly, Mistral’s stated strengths in agentic coding and specialized professional tasks are company claims that require testing on relevant workloads.

The benchmark table is a dated snapshot, and its settings differ: some scores use maximum reasoning effort, while others use high effort or a default fallback. The comparison therefore offers a broad signal, not a controlled head-to-head trial. Developer location in the table refers to where a company is based, not where an individual API request is processed.

“The company said it trained Large 4 on its own infrastructure in Europe and continues to improve it.”

— Mistral

Preview Performance Still May Change

Several limits remain. The weights had not been released at the time of the source report, and Mistral said it was still improving the model. It is not clear whether the later weight release will include a materially different version, or how that version will perform against the same benchmark suite.

The available index score does not establish how Large 4 performs on a particular user’s tasks, how reliably it completes long agentic workflows, or how often it produces unsupported claims. The source author’s hallucination observations are anecdotal. The supplied source material also cuts off during its discussion of cost, so it does not provide enough information to make a supported price comparison. Benchmark scores and reasoning settings may also change over time.

Weight Release and Workload Tests

Mistral’s next stated milestone is a release of Large 4’s weights later in October. Until that happens, developers can evaluate the API preview on their own tasks, while treating current scores as a dated snapshot. A meaningful decision will require testing the same workflows across candidate models, including how often they follow constraints, verify results and need human correction.

Further Artificial Analysis evaluations or updated model profiles may change the benchmark comparison. For now, the source report’s conclusion is limited: Large 4 shows progress for Mistral, but the cited evidence does not put this preview on par with the higher-scoring models for demanding, extended work.

Key Questions

How does Mistral Large 4 score against the models listed?

It scored 38 on Artificial Analysis’s Intelligence Index in the October 7, 2026 snapshot. That was below the listed scores for Claude Opus 5.5, Gemini 4 Argon, GPT-6.1 Sol, GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. It matched GPT-6 Luna in the cited settings and scored above Cohere Command A+.

Can developers download Large 4’s weights now?

Not according to the source report, dated October 7, 2026. Mistral had made the model available as a public API preview and scheduled a weight release for later in October.

Does a score of 38 prove the model will fail at agentic tasks?

No. The index is an aggregate benchmark, not a direct prediction of success or reliability on every workflow. The source author uses the score, alongside personal experience, to explain a cautious model-selection judgment; it is not proof that Large 4 will fail a specific task.

Is the report’s hallucination comparison a controlled study?

No. The author says he encountered hallucinations while using the preview, but describes that as personal experience, not a controlled comparison measuring hallucination rates across models.

Source: ThorstenMeyerAI.com