🔍 Read the full analysis: Mistral Large 4: Why It Stands Out Abroad But May Not Power Your Agents on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Mistral released Large 4 as a research preview, scoring 38.4 on the Artificial Analysis Intelligence Index. That is a major improvement over its earlier models and makes it a leading model from outside the United States and China, but it trails major US and Chinese competitors; benchmark cost and reported verbosity also raise questions about using it for agents.
Mistral released Large 4 as a Research Public Preview, and the model scored 38.4 on Artificial Analysis Intelligence Index v4.3.2. The result marks a steep improvement over Mistral’s earlier models, but leaves Large 4 behind leading US and Chinese systems on the same index, which tests work that includes agentic tasks.
Artificial Analysis lists Large 4 at 38.4 points, compared with 9 for Mistral Large 3 and 14 for Medium 3.5 on the same index version. The source describes that change as the largest step taken in one release by a European lab. The score still trails the highest-ranked models in the listed comparison: Claude Opus 5.5 is at 57.6, while several Chinese models also score above Large 4, including GLM-5.3 at 44.8 and DeepSeek V4.1 Flash at 39.5.
Mistral describes Large 4 as a one-trillion-parameter model with 49 billion active parameters, text-and-image input, text output and a 512,000-token context window. It is currently offered as a proprietary API preview. Mistral has said the model weights are expected at the end of October, but the licence has not been published in the supplied material. Standard API pricing is listed as $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14; the source reports a 50% discount for the first two weeks.
Artificial Analysis data cited by the source put Large 4’s cost at $1.13 per Intelligence Index task. GLM-5.3-Flash is listed at $0.25 per task with a score of 41.8, while DeepSeek V4.1 Flash costs $0.27 and scores 39.5. These comparisons describe benchmark task costs, not a universal price for every workload. The source also says Large 4 generated 200 million output tokens across the index, against a median of 81 million for comparable models; that figure may affect costs in tasks that require similarly long outputs.
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.
The Trade-Off for Agent Workloads
The result matters to businesses choosing a model for multi-step agents, not just for ordinary chat. According to the source, Artificial Analysis’s index includes agentic knowledge work, real-world work tasks, software workflows and coding benchmarks. A model’s overall score is not a direct pass rate for a specific company workflow, but weaker performance across tasks can raise the risk that an agent makes an error and carries it into later steps.
The cost comparison adds a separate procurement concern. Large 4’s reported benchmark task cost is higher than those of two models that score above it on the index. The source also reports unusually high output volume, which can increase latency and spending when a system pays for generated tokens. These figures do not prove Large 4 will be more expensive for every use case: actual costs depend on prompts, output lengths, caching and task design.
For European organisations seeking a model from outside the US and China, Large 4 offers a notable option and a sizeable improvement over Mistral’s prior scores. But the supplied comparisons do not support treating it as a peer to the leading frontier systems. The sensible takeaway is that its suitability depends on measured performance and cost in the intended application, rather than the geographic framing of its launch.
Popular right now
Dyson Ball Animal 3 Corded Upright Vacuum
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
KitchenAid Artisan 5-Quart Tilt-Head Stand Mixer, Almond Cream
As an affiliate, we earn on qualifying purchases.
How Large 4 Compares
The source’s comparison uses Artificial Analysis Intelligence Index v4.3.2, presented as the current index version. It lists six leading US models above 50 points, with the top score at 57.6. Large 4’s 38.4 sits below multiple Chinese models, including GLM-5.3, Kimi K3 at 43.6 and GLM-5.3-Flash at 41.8. It is above some older or lower-scoring entries, including GLM-5.2 at 33.7 and DeepSeek V4 Pro at 36.0.
The source characterises Large 4 as the highest-scoring model outside the US and China in its comparison. That should be read as a narrow description of the models included, not proof that it competes with the overall leaders. The same source says Mistral has not yet released the model weights and that the licence is unpublished, so its current preview access should not be confused with an already available open-weight release.
“Reinforcement learning is still running.”
— Mistral, as reported in the supplied source
Preview Limits and Open Questions
Large 4 is still a research preview, and Mistral says reinforcement learning is ongoing. The final benchmark results may differ, while the source does not provide a date for an updated evaluation. The weights’ anticipated end-of-October release and their eventual licence terms also remain unconfirmed in the material provided.
The source author’s report of confident hallucinations is an anecdotal hands-on observation. It is not quantified or tied to a published test protocol, so it cannot establish how often Large 4 produces false claims. Likewise, benchmark task cost and output volume may not predict spending or reliability in an individual deployment. No details are supplied about the prompts, repeated-run variability, or an independent agent workflow evaluation specific to Large 4.
Weights and Updated Evaluations
The next stated milestone is Mistral’s planned release of Large 4’s weights at the end of October. Buyers will need to check the published licence and deployment terms when that release arrives; neither is specified in the supplied source. Mistral may also update the model after its ongoing reinforcement-learning work.
For now, organisations considering the API preview can compare Large 4 against alternatives on their own tasks, recording completion rates, error propagation, output length, latency and total cost. Further Artificial Analysis results or reproducible evaluations would help establish whether the current ranking changes and how well the model performs in practical agent workflows.
Key Questions
What is Mistral Large 4?
Mistral Large 4 is Mistral’s newly released research-preview model. The supplied source describes it as a one-trillion-parameter model with 49 billion active parameters, text and image input, text output, and a 512,000-token context window.
How did Large 4 score?
It scored 38.4 on Artificial Analysis Intelligence Index v4.3.2, according to the supplied source. That is above the listed scores for Mistral Large 3 and Medium 3.5, but below several leading US and Chinese models.
Is Large 4 open weight now?
No. The source says it is currently a proprietary API preview. Mistral has said weights are planned for the end of October, but the licence has not been published in the supplied material.
Why might it be a concern for agents?
The source points to its lower index score than several rivals, higher reported benchmark task cost than two Chinese models that outscore it, and high output-token volume. These are reasons to test it on a specific workflow, not proof that it will fail or cost more in every deployment.
Are the reported hallucinations independently confirmed?
No. The hallucination concern in the source is the author’s hands-on observation. The material does not provide a quantified independent test of Large 4’s hallucination rate.
Source: ThorstenMeyerAI.com
