Mistral Large 4 Review: Price, Benchmarks, and Verdict

This Mistral Large 4 review examines Mistral AI's new trillion-parameter, multimodal model, released in public preview on October 6, 2026. The model combines aggressive introductory API pricing with a large advertised context window, strong specialist benchmark results, and a promised open-weight release later in October.

The launch needs careful qualification. The weights and final license are not yet available, independent tests do not make Mistral Large 4 the strongest general-purpose model, and published context figures differ across providers and evaluators. This is an evidence-based assessment of current documentation and third-party results, not a claim of hands-on testing.

Mistral Large 4 Review: Quick Verdict

Mistral Large 4 deserves a place on evaluation shortlists for document-heavy work, legal and financial agents, multimodal analysis, and organizations that want a path toward self-hosting a European model. Its current promotional API rates—$0.68 per million input tokens and $2.09 per million output tokens—make a pilot relatively inexpensive compared with many proprietary frontier models.

The evidence does not support replacing an entire model stack on launch day. Vals AI places the preview 32nd of 44 models on its broad index, although it ranks sixth on Harvey's narrower Legal Agent Benchmark. Artificial Analysis finds above-average intelligence and fast output, but also unusually high token use. Early DaBench results are weaker still on its small set of code-graded real-world tasks.

The practical verdict is workload-specific: test Mistral Large 4 where its open-weight roadmap, price, long-document performance, or specialist results matter. Keep a stronger generalist available, record the exact preview version, and wait for the downloadable weights and license before treating self-hosting as a current capability.

What Is Mistral Large 4?

Mistral Large 4, nicknamed “Le Chonk,” is Mistral's largest model so far. Its granular mixture-of-experts architecture contains about 1.05 trillion total parameters, with 49 billion active expert parameters per token. Mistral's current model card reports 52 billion active parameters when embeddings and output layers are included, plus a 1.6 billion-parameter vision encoder.

The model accepts text and images and returns text. Mistral lists structured outputs, function calling, document question answering, prefix completion, batch processing, built-in tools, and its Agents and Conversations APIs. That feature set makes it relevant to coding, research, document analysis, and tool-using workflows rather than only chat.

Mistral currently offers mistral-large-4 through its API in public preview. The company says weights will arrive at the end of October, while its official Hugging Face countdown lists October 31. Until the artifacts and license are published, “open-weight” describes the announced release path—not a model that teams can already download and deploy.

Mistral Large 4 Specs and Context Limits

Mistral's official card lists a one-million-token context window. Vals AI currently records 512,000 tokens, while Artificial Analysis reports 524,000. These figures may describe different preview configurations, combined input-output limits, or provider metadata, but the discrepancy is unresolved in the public material reviewed for this article.

Do not design a production pipeline around the largest number without testing the exact endpoint. A context window is capacity, not a guarantee that the model will retrieve every relevant fact from a long document set. Evaluate citation accuracy, missed evidence, ordering effects, latency, and cost at the lengths your application actually uses.

The same version discipline applies to parameter counts. “49B active” and “52B active” are not necessarily contradictory: Mistral explains that the higher figure includes embeddings and output layers. Neither number indicates how much memory a self-hosted deployment will require, because all trillion-plus stored parameters still affect weight storage and serving architecture.

Mistral Large 4 Benchmarks: What the Results Show

Mistral's launch post highlights agentic coding and specialist enterprise work. It reports 61.7% on DeepSWE 1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.0, producing a 49.8% Coding Agent Index. A Mistral-commissioned blind human evaluation ranked the preview second among five tested models for coding quality, behind Claude Opus 5.

Those are useful signals, not neutral proof of universal leadership. Vals AI reports a different Terminal-Bench 4.0 result of 22.73%, using its published configuration. Harness, task selection, execution budget, and scoring can materially change agent results, so figures from separate evaluators should not be averaged or compared as if they came from one run.

The specialist evidence is more interesting. Vals gives Mistral Large 4 a 54.68% score on Finance Agent v2 and 15.83% on Harvey's Legal Agent Benchmark. The legal result ranks sixth of 75 models in that snapshot, but its low absolute pass rate still argues for qualified review. On Vals' broader index, the preview scores 48.05% and ranks 32nd of 44 models, behind several closed and open alternatives.

Artificial Analysis gives the preview an Intelligence Index score of 38, measures about 116 output tokens per second, and describes it as fast but very verbose. Its current pricing calculation uses the original $1.36 input and $4.18 output rates, not Mistral's lower displayed promotional rates. DaBench's early code-graded suite places it ninth with an overall score of 28. These snapshots are likely to move as the preview and evaluation coverage mature.

Mistral Large 4 Pricing and Real Cost

Mistral's current model card shows sale prices of $0.68 per million input tokens, $0.07 per million cached input tokens, and $2.09 per million output tokens. Crossed-out prices are twice as high: $1.36, $0.14, and $4.18. Because this is a public preview with explicitly promotional pricing, verify the live card before budgeting or comparing providers.

Low token rates do not guarantee low task cost. Reasoning length, verbose output, retries, tool calls, long context, caching, and agent loops can multiply consumption. Artificial Analysis generated substantially more output tokens than its median model while evaluating the preview. A model that produces a cheap first answer but needs several correction turns can cost more than a higher-priced model that passes the acceptance test immediately.

Measure cost per accepted result. For each representative task, record input, cached input, reasoning and answer tokens, latency, retry rate, reviewer time, and failure cost. Compare the same prompts, tools, output schema, and stopping conditions. SD's Claude Haiku 5.5 review and GLM-5.3-Flash review provide useful reference points for current low-cost model evaluation.

Open Weights, Self-Hosting, and Sovereignty

The planned weight release is Mistral Large 4's clearest strategic difference from closed frontier models. Downloadable weights can support private infrastructure, auditing, fine-tuning, and deployment choices that do not depend on a single hosted API. Mistral also emphasizes training and serving infrastructure in Europe.

That promise is not yet a completed deployment option. The official Hugging Face page still shows an upcoming release, and no final license or full serving guide was available during this review. A one-trillion-parameter mixture-of-experts model is also not lightweight: only a fraction of parameters may be active for each token, but storage, networking, memory placement, quantization, and parallel inference still create substantial operational demands.

Open weights do not remove security or compliance work. Teams remain responsible for access controls, data handling, model and dependency provenance, output monitoring, and safe tool execution. For agent deployments, SD's AI agent security guide explains why least privilege, sandboxing, approval gates, and audit trails still matter.

Best Use Cases and Limitations

The most credible early use cases are long-document analysis, legal and financial research with expert review, structured extraction, multilingual and multimodal knowledge work, and agent workflows where current API price matters. Organizations evaluating European infrastructure or future private deployment have an additional reason to test it.

Mistral Large 4 is a weaker default choice when a team needs the strongest broad benchmark performance, a stable generally available API, a published self-hosting license today, or predictable short outputs. The unresolved context metadata and fast-moving preview pricing add implementation risk. Early benchmark disagreement also shows why model choice cannot be reduced to one launch chart.

For end-user work and coding, the model should be evaluated separately from the product harness. Mistral's Vibe agent adds connected tools, permissions, sandboxes, and subscription limits that are not measured by base-model benchmarks. A good model inside a weak workflow can still fail, while a strong harness can improve a less capable model through tools, verification, and human review.

Conclusion

This Mistral Large 4 review finds a promising but unfinished frontier-model release. Aggressive preview pricing, fast output, multimodal support, specialist legal and finance results, and a planned open-weight release give it a distinct case for evaluation. Its European infrastructure story may also matter to organizations seeking more deployment control.

The caveats are material: broad independent results are mixed, output can be verbose, context metadata conflicts, pricing is promotional, and the weights and license are still pending. Test Mistral Large 4 on the workflow you actually operate, keep human review for high-impact domains, and make a production decision only after the final artifacts, terms, and stable endpoint are available.