Mistral Large 4 cuts benchmark task cost 63% after its preview

A cheaper release, not a new model family

Mistral had offered Large 4 through its API and promised to release the weights on October 27. They arrived two days later. The release reflects additional training and deployment options, not a separate model family.

Large 4 is a sparse-activation model that takes images and text as input and generates text. Artificial Analysis lists a 1 million-token context window. Mistral’s release materials give it 1.05 trillion total parameters and 52 billion active parameters.

The active-parameter count differs from the 49 billion reported in VentureBeat’s earlier coverage. Mistral says the smaller figure counts activity per token, while the larger also includes embeddings and output layers. That discrepancy alone does not establish that the model grew during the preview period.

Mistral also revised its reported training hardware from about 4,000 GPUs to 3,800 Nvidia Grace Blackwell GPUs. The weights can be run with vLLM or SGLang. Mistral lists deployment through its Studio and API, OpenRouter, Vibe Code, Microsoft Foundry, Nvidia endpoints and an optimized NIM container; Microsoft Foundry also supports private deployment in Azure.

The release offers three reasoning modes: low, medium and high, with high as the default for difficult requests. Developers can use deeper reasoning for hard debugging and a lower setting for routine queries. Whether that reduces the cost of completed work depends on how often a cheaper attempt needs correction.

The cost drop has limits

For the released model, Mistral set API prices at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Artificial Analysis confirmed those prices. The preview rates were $1.36, $0.14 and $4.18, respectively—roughly twice as much.

Artificial Analysis now measures task costs for the released model rather than estimating them from token prices. Its October 29 estimate puts Large 4’s Intelligence Index task cost at $2.06, down from $5.54 for the preview. The model page reports the same figure. These numbers refer to Intelligence Index tasks, not the Coding Agent Index.

That brings Large 4 close to Muse Spark Open on cost, but DeepSeek and MiMo score higher on Intelligence Index while costing substantially less per task. Output-token use alone does not determine the bill: input volume, caching and token rates matter too.

Nor does the benchmark cost include self-hosting, configuration, operational oversight or human review. The relevant figure for a company is the cost of work completed to an acceptable standard in its own environment—not a benchmark task in isolation.

There is also a business constraint attached to the open weights. Mistral’s modified MIT license permits running, copying, modifying, fine-tuning and distributing the model, including derivatives, subject to notice requirements. But it adds a commercial restriction: under Section 2, a separate commercial license is required before commercial use if the licensee or an affiliate provides models as a service.

The definition covers giving third parties access to inference or fine-tuning through an API, hosted endpoint or other means that lets users substantially control inputs, parameters or training data. It excludes certain user-facing products where model capabilities appear only through specific features or software wrappers, as well as services that merely route requests to models hosted by other companies. Section 3 separately excludes internal use when the model or derivatives are not made available to third parties, and use through Mistral’s official products.

In practice, the dividing line is how a company provides access, not its size: the license sets no revenue threshold. Permission to run downloaded weights does not mean hosted services or self-hosting infrastructure are free.

A strong coding result depends on the setup

Artificial Analysis placed Large 4 second among open-weight models available at the time in its Coding Agent Index. It trails GLM-5.3 by one point and leads DeepSeek by two. Large 4 uses 8.5 million tokens per task, about 47% fewer than GLM’s 15.9 million.

That is evidence of efficiency, but the index evaluates full agent configurations: the model plus the software that manages tool calls and coding tasks. It does not show how the models would compare if every other component were held constant. The Coding Agent Index includes FrontierCode v1.1, Terminal-Bench 4.0 and SWE-Atlas-QnA-AA; it is separate from the broader Intelligence Index.

The benchmark differences matter. Mistral reports that Large 4 scored 41% on Terminal-Bench, up from 28% for the preview, while using 30% fewer output tokens. Artificial Analysis reports 25% for the same test in its Intelligence Index. The results come from different conditions, and the available materials do not isolate how much software and settings explain the gap. Neither figure is a universal measure of the model.

FrontierCode tells a similar story. Mistral’s chart gives Large 4 19.7%, against DeepSeek’s 19.3%. Artificial Analysis scores Large 4 at 15%, behind DeepSeek at 29% and GLM-5.3 at 22%. The independent result does not support Mistral’s lead outside its own comparison.

Mistral also cites a blind preference study by Surge Code. Large 4 won 64% of comparisons with GLM-5.3 and Kimi K3, 56% with DeepSeek V4.1 Flash and 22% with Claude Opus 5.5. Those are human preferences in the stated matchups, not a substitute for task benchmarks.

Mistral’s earlier preview coverage focused on a 62% result in DeepSWE. The new materials do not provide an updated DeepSWE score, and neither Terminal-Bench nor the composite coding ranking can be directly compared with it.

The open-model lead is narrowly defined

Large 4’s final Intelligence Index score of 26 is comparable with the preview’s 25 because both use version 5 of the benchmark. That is a small gain, not evidence of a sharp jump in capability. Artificial Analysis found more noticeable changes in token use and hallucination rate.

Large 4 is one point ahead of Muse Spark Open, making it the highest-scoring open-weight model developed outside China in this evaluation. It is not the world’s strongest open model: seven open models score higher. DeepSeek V4.1 Flash 1017 and GLM-5.3 each score 32, while MiMo-V2.6-Pro scores 30. Among all models in the ranking, Gemini 4 Carbon scores 59, Claude Opus 5.5 scores 58 and GPT-6 Astra scores 57. These are high-reasoning configurations.

Artificial Analysis attributes much of Large 4’s gap with the open-model leaders to coding and knowledge useful to agents. On AA-Briefcase, Large 4 scores 1,349 Elo, against 1,570 for DeepSeek and 1,504 for GLM. Its clearest edge over Muse Spark Open is Terminal-Bench, where it scores 25% against Muse’s 8% in Intelligence Index.

The efficiency improvement is more pronounced. Artificial Analysis says the final version uses an average of 82,000 output tokens per Intelligence Index task, down 39% from 135,000 for the preview. The seven higher-scoring open models use between 113,000 and 225,000 output tokens per task. That combination of score and output use does not make Large 4 the cheapest option: these averages exclude input tokens, and the coding index’s 8.5 million tokens per task includes input, cached and output tokens across a different set of tasks.

Earlier scores also need care. Artificial Analysis gave the preview 38 points when it launched in October. Intelligence Index v5 later added harder tasks and raised the share of tests with closed or hidden answers to 60%. The old 38 and the current 26 use different scales; the like-for-like comparison is 25 to 26.

I think the more useful question is not whether Large 4 leads a ranking, but whether the cost advantage survives in a company’s own workload. Mistral says existing customers have seen their bills fall over the past six months, but its spokesperson said systems using Large 4 are still being built and it is too early to measure meaningful results in production. The release offers weights and deployment choices; it does not yet show what customers save on completed work.