🔍 Read the full analysis: Why Mistral Large 4 Hasn’t Caught Up In The AI Race on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Mistral released Large 4 in public preview on October 6, 2026, with open weights scheduled for later in the month. Artificial Analysis gave the preview an Intelligence Index score of 38, below several leading U.S. and Chinese models; a reviewer also reported hallucinations in personal use. Those findings raise questions about its suitability for demanding agentic tasks, but they do not establish how it will perform across all workloads or after further updates.
Mistral Large 4 entered public preview on October 6, but the model’s initial benchmark results put it behind several leading U.S. and Chinese competitors, leaving its standing in complex AI work unsettled. Artificial Analysis scored the preview 38 on its Intelligence Index, while a reviewer at ThorstenMeyerAI.com said personal experience with hallucinations made them reluctant to use it for demanding, long-running agentic tasks.
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. It accepts text and images. The release currently available is a preview API; the model weights are scheduled for release later in October and were not publicly downloadable as of October 7. Mistral says the model was trained on its own infrastructure in Europe and remains under development.
In the Artificial Analysis comparison cited by the reviewer, Large 4’s score of 38 matched OpenAI’s GPT-6 Luna at maximum reasoning effort and was just below DeepSeek V4.1 Flash at maximum effort, which scored 39. Other listed scores were 58 for Anthropic’s Claude Opus 5.5, 53 for Google’s Gemini 4 Argon, and 52 for OpenAI’s GPT-6.1 Sol. China’s GLM-5.3 scored 45 and Kimi K3 scored 44. These are dated scores from different named reasoning settings, not tests run under identical compute budgets.
The reviewer’s conclusion is narrower than a claim that Large 4 cannot do useful work: they would not choose the current preview for demanding agentic work or long tasks when higher-scoring alternatives are available. The review also reports hallucinations during personal use. That observation is an individual account, not a controlled comparison of hallucination rates across models. Artificial Analysis reports about 512,000 tokens of context capacity, but context size measures how much material can be supplied, not whether the model reasons accurately over it.
Benchmark Gaps Matter for Agentic Work
For developers choosing a model to plan, use tools and carry decisions through a multistep task, aggregate benchmark performance can help narrow the options. A weak assumption made early in a workflow may shape later actions, while a polished final answer can make mistakes harder to spot. The reviewer argues that sustained execution and evidence-checking matter more than headline specifications when a model is expected to work with limited supervision.
The score is relevant evidence, but it does not predict success on every coding, research or business task. Nor does a 20-point gap between Large 4 and Claude Opus 5.5 mean a 20% difference in intelligence or task success: these are index points, not percentages. Mistral says the preview has strengths in agentic coding and specialized professional tasks; those claims still need testing on the specific workloads developers care about.
The launch also matters beyond model rankings. Mistral says it trained Large 4 on European infrastructure, a development in the region’s capacity to build large AI systems. That is distinct from proving the preview is the best choice for a given application. The evaluation question for users remains whether its quality, reliability and cost fit their work.
As an affiliate, we earn on qualifying purchases.
A Preview, Not the Final Release
The comparison is a snapshot dated October 7, 2026, based on Artificial Analysis scores cited in the review. Scores and model offerings can change, and the named reasoning settings differ. The source also cautions that developer locations identify the organizations behind the models, not where an individual API request is processed.
The review places Mistral behind several U.S. frontier models and the listed GLM-5.3 and Kimi K3 models, while noting that Cohere’s Command A+ scored 13. That counterexample means the comparison does not support saying every competing lab is ahead. The source says DeepSeek V4.1 Flash has approximately comparable benchmark intelligence at a much lower measured cost per task, but the supplied material ends before giving the relevant cost figures or methodology.
Mistral’s weights were scheduled for a later October release, and the company says it is continuing to improve the model. The present judgment therefore concerns an API preview, not a completed assessment of a publicly downloadable final version.
“I would not choose it for demanding agentic work or long tasks when stronger models are available.”
— Thorsten Meyer, ThorstenMeyerAI.com
AI hallucination detection software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Reliability and Costs Still Need Testing
It is not yet clear how Large 4 will perform on independent, workload-specific evaluations, including extended coding and research tasks. The Intelligence Index is an aggregate measure, not a direct test of reliability on the reviewer’s workflows. The reported hallucinations reflect one person’s experience; the source provides no controlled rate or head-to-head study.
The final weights had not been released as of October 7, and future model updates could change results. The supplied source also does not include the full cost comparison it references, so readers cannot verify the exact price difference or the conditions behind the claim that DeepSeek V4.1 Flash costs less per task. API pricing and performance under comparable settings remain open questions.
As an affiliate, we earn on qualifying purchases.
Watch for Weights and Independent Tests
Mistral has scheduled release of Large 4’s weights for later in October 2026, though the source gives no precise date. The company says it is still improving the model. After the weights become available, developers and independent evaluators can assess its behavior outside the preview API and test it against specific coding, research and professional workloads.
For now, teams considering the preview should compare it on their own tasks, including how often outputs require correction and how much supervision multistep work needs. New benchmark results, clearer pricing details and evaluations of the eventual weight release will show whether the current performance gap narrows.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Mistral Large 4’s current benchmark score?
Artificial Analysis gave the preview a score of 38 on its Intelligence Index in the comparison cited on October 7, 2026. The score is an aggregate benchmark result, not a percentage or a guarantee of performance on an individual task.
Are Mistral Large 4’s weights available to download?
Not as of October 7, 2026. Mistral had launched a preview API and scheduled the weights for release later in October. The source does not give an exact release date.
Does the score prove Large 4 is unreliable?
No. The score places it below several listed competitors on that benchmark, but it does not prove the model will fail a particular task. The reviewer’s report of hallucinations is personal experience, not a controlled comparison of error rates.
Why is the model being questioned for agentic work?
Agentic work involves multiple steps, tool use and decisions that can depend on earlier results. The reviewer says the benchmark gap and personal experience with hallucinations are reasons to prefer stronger alternatives for demanding, long tasks, while acknowledging that workload-specific tests are still needed.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
