🔍 Read the full analysis: The Case For Mistral Large 4 Outside The US And China—and Its Agent Limits on ThorstenMeyerAI.com
Get hardware and tech essentials 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
TL;DR
Mistral Large 4 recorded 38.4 on Artificial Analysis’s Intelligence Index v4.3.2, a sharp rise from its predecessor and a strong result among models outside the United States and China. The source’s benchmark and pricing comparisons suggest it trails leading US and Chinese models, while concerns about verbosity and observed hallucinations may limit its suitability for long-running agents. Its weights, licence and final benchmark performance remain unsettled.
Mistral describes Large 4 as a **one-trillion-parameter model with 49 billion active parameters**, capable of taking text and images as input and producing text. It has a **512,000-token context window** and is available through Mistral’s API as a research public preview. The source says Mistral plans to release the model weights at the end of October; until then, it is proprietary, and the licence has not been published.
Artificial Analysis’s current Index puts Large 4 at **38.4 points**, up from 9 for Mistral Large 3 on the same Index version. That is a substantial improvement, but the cited table ranks it below the leading US models and several Chinese models, including GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. The source characterizes the result as a European advance, not evidence that Mistral has reached the top tier.
The listed API price is **$1.36 per million input tokens and $4.18 per million output tokens**, with cached input at $0.14 per million. The source reports a 50% introductory discount for the first two weeks. It also says Artificial Analysis measured 200 million output tokens for Large 4’s Index tasks, compared with an 81 million median for comparable models, making output volume relevant to the overall cost of benchmark work.
Mistral Large 4: best outside the US and China — and still not a model to run your agents on
The headline is true: France has the most intelligent model outside the US and China. The independent data says the rest: every US and Chinese flagship scores higher, the best by 19 points. It costs 4× more per task than Chinese open models that outscore it, and it’s 2.5× as verbose as the median model.
~two-thirds of Opus 5.5. Level with OpenAI’s small model, Luna.
Eighth among open models once weights ship — behind seven Chinese ones. Beats GLM-5.2 and V4 Pro, loses to their successors.
Cohere doesn’t compete at this tier — reported ~14% hallucination at ~9% accuracy, because it declines most questions. A field of one.
The Index is now agentic-heavy — Briefcase, GDPval, AutomationBench, Terminal-Bench. Errors multiply across steps: tolerable in chat, fatal over a two-hour run.
AA v4.3.2Output tokens to complete the Index. On an agent, verbosity is cost and latency on every step.
AAConfident false assertions in hands-on use. US frontier has largely moved past this — Gemini 4 Argon: 15%. In fairness Chinese open models are worse (Kimi K3 51%, DeepSeek V4 Pro 94%). In an agent, a fabrication is a wrong premise every later step builds on.
AUTHOR’S TESTING · not an AA figure- Cyber defence: 50 on the AA Cyber Index; 82% CyberGym-E2E (ahead of Luna’s 78%). Likely top-3 open model on cyber.
- Documents & images: 19% GDP.pdf (+18 vs Large 3); 100 images per request.
- Speed: 116 tok/s, 1.46s TTFT — well above median.
- The jump: Large 3 scored 9 on this Index. 9 → 38 is real progress.
- Jurisdiction: French parent, EU hosting, weights promised end of October.
- Legally bound buyers (defence, classified, DORA, health data): now the best European option by a wide margin. Wait for the weights, check the licence, pilot on cyber and documents.
- Everyone else, for agentic or long tasks: don’t. A US frontier model is meaningfully more capable; GLM-5.3-Flash is more capable and 4× cheaper.
- Note: Preview — Mistral says RL is still running, so scores may move. That changes next month’s decision, not today’s.
Mistral says it has “essentially closed the gap.” It has closed the gap to where the Chinese open-weights field was a few months ago, while that field and the US frontier have both moved on. On every independent measure that matters for agents — intelligence, cost per task, verbosity and factual reliability — Large 4 is not a frontier model. “Most intelligent outside the US and China” is true mainly because almost nobody else outside those two countries is competing. Use it if you have to. Don’t use it because of the headline.
Benchmark Gains, Agent Trade-offs
The result matters to organizations looking for a capable model from a European provider, particularly those seeking alternatives to US and Chinese suppliers. Large 4’s **jump from 9 to 38.4** marks a clear improvement in Mistral’s measured performance. But the source’s comparisons make clear that geography alone does not establish technical or economic advantage: several Chinese models score higher and are reported to cost much less per benchmark task.
For agent use, a model’s ability to sustain multi-step work matters alongside its headline score. The source argues that **errors can compound across a long workflow**, while verbosity adds output cost and latency at each step. Its report of confident hallucinations comes from hands-on testing, not the Artificial Analysis benchmark, and should be treated as an attributed observation rather than an independently established rate. Still, it highlights a practical risk: an incorrect answer used as an agent’s premise can affect later actions.
These findings do not establish that Large 4 is unsuitable for every deployment. They point instead to a need for task-specific testing, including reliability, token use, latency and total cost. For high-stakes or extended workflows, benchmark rank alone may not capture the operational risk.
As an affiliate, we earn on qualifying purchases.
A Sharp Rise From Large 3
The comparison in the source uses **Artificial Analysis Intelligence Index v4.3.2**, which includes evaluations of agentic knowledge work, real-world work tasks, software workflows and coding. That mix makes the score relevant to agent capabilities, though it remains a benchmark result rather than a guarantee of performance in a particular company’s tools or tasks.
Mistral’s earlier Large 3 scored **9** on the same Index version, while Medium 3.5 scored 14. Large 4’s result is therefore a major step within the company’s own lineup. The source also notes that Mistral says reinforcement learning is still underway and that scores may change, so the preview result should not be treated as necessarily final.
The phrase “most intelligent model outside the US and China” depends on which competitors are included. The source says few labs in other regions are competing at this level and describes the comparison as a field with limited entrants. That geographic distinction may matter for procurement, but it is separate from how Large 4 performs against the strongest models overall.
“Reinforcement learning is still running, so scores may move.”
— Mistral
large language model with 512k token window
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Preview Results Still May Change
Large 4 is a **research preview**, and Mistral says reinforcement learning is ongoing. The source provides no final benchmark result, and the announced end-of-October timing for weights is a plan rather than a completed release. The model’s licence is also unpublished, leaving a material question for organizations considering self-hosting or redistribution.
The source does not provide a controlled, independently verified hallucination rate for Large 4. Its account of confident errors is explicitly based on hands-on use. Nor does the benchmark alone show how the model will perform on a buyer’s specific agent workflow, where tools, prompts, safeguards and task length can change results.
Pricing comparisons also depend on usage patterns. The source gives API rates and per-task benchmark estimates, but those figures do not settle total deployment cost across different input lengths, caching, retries or agent designs. The introductory discount is time-limited, and the source does not specify its end date beyond saying it lasts two weeks.
As an affiliate, we earn on qualifying purchases.
Weights, Licensing and Retesting
The next concrete milestone is Mistral’s planned **release of Large 4’s weights at the end of October**. Buyers will be able to assess the licence once it is published and determine whether the release permits their intended uses. Until then, access described in the source is through Mistral’s API preview.
Artificial Analysis scores may also change as Mistral’s reinforcement learning continues. Organizations evaluating the model should compare updated results with their own tests, particularly for long-running agents, and track output-token use alongside accuracy and latency. The source does not identify a confirmed date for a final benchmark or a production release.
As an affiliate, we earn on qualifying purchases.
Key Questions
What is Mistral Large 4?
It is Mistral’s **one-trillion-parameter model**, with 49 billion active parameters, text and image input, text output and a 512,000-token context window. The source describes it as available through the API in research preview.
How does Large 4 score against other models?
Artificial Analysis’s Index v4.3.2 gives it **38.4 points**. The cited comparison places it below leading US models and several Chinese models, while above Mistral Large 3’s score of 9 on the same Index version.
Is Large 4 suitable for AI agents?
The available information does not establish a universal answer. The source flags verbosity and reports observed confident hallucinations, concerns that may matter in multi-step work. Teams should test the preview on their own tasks before relying on it.
When will the model weights be available?
Mistral’s reported plan is to release the weights **at the end of October**. The source says the licence has not yet been published, so the terms of use remain unknown.
How much does API access cost?
The reported standard price is **$1.36 per million input tokens** and **$4.18 per million output tokens**, with cached input at $0.14 per million. The source reports a 50% discount for the first two weeks; actual costs depend on usage. Model outputs and API use can carry financial and operational risk, and these figures are not a recommendation to use or avoid the service.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
