Mistral releases Large 4 in preview and expands Studio platform

Mistral AI launched Mistral Large 4 (Le Chonk) and expanded its Studio enterprise platform to help businesses build, deploy and govern AI agents, workflows and applications across multiple environments.

By Mira Sethi

November 5, 2026 at 11:03 AM EST2 min read

Key takeaways (3)

1.Artificial Analysis scored the Mistral Large 4 Preview 38 on its Intelligence Index and 50 on its Cyber Index.
2.Artificial Analysis put standard cost at $1.13 per task, or $0.57 with a two-week discount, above two cheaper rivals.
3.Customers named by Mistral use Studio in production; incident, routing and staffing examples are Mistral's internal claims.

Key takeaways (3)

1.Artificial Analysis scored the Mistral Large 4 Preview 38 on its Intelligence Index and 50 on its Cyber Index.
2.Artificial Analysis put standard cost at $1.13 per task, or $0.57 with a two-week discount, above two cheaper rivals.
3.Customers named by Mistral use Studio in production; incident, routing and staffing examples are Mistral's internal claims.

Mistral AI announced Mistral Large 4, codenamed Le Chonk, and an expansion of its Studio enterprise platform in a company post dated 5 November.

The model is in Research Public Preview on Mistral's API, according to Artificial Analysis 2, which scored it 38 on its Intelligence Index. The same firm puts its cost per task at more than four times that of open-weights models of similar intelligence.

What Mistral says

Mistral said Large 4 "does great for coding, delivers advanced defensive cybersecurity capabilities, and brings powerful reasoning to complex documents and business tasks," according to its announcement 1. The post makes no comparison with earlier models and gives no benchmark figures.

The post also says Studio gains managed agents and apps 1, along with new capabilities for workflows, enterprise context, agent observability and evaluations. Event triggers from enterprise systems and a visual graph of workflow logic are also added. Mistral said Studio can run in the cloud, in a customer's VPC, on-premises or across a combination, with human approval required for actions the customer chooses.

Mistral said ABANCA, Accenture, Alten, BNP Paribas, Capgemini, CMA CGM and Tata Consultancy Services already use Studio in production 1. The post names no tasks for those customers.

The three worked examples in the post come from Mistral's own teams. The company said an incident-investigation agent built by its engineering team saves "hundreds of hours each month," and a workflow built by its product operations team saves about 35 hours a week. Its Applied AI team built a staffing app. These are Mistral's figures for internal use, and the post does not describe how they were measured.

What Artificial Analysis measured

Artificial Analysis tested the Mistral Large 4 Preview, a 1T-parameter model with 49B active parameters that takes text and image input and produces text. These are its own results, and its article does not describe replication by others.

The preview scored 38 on the Intelligence Index, comparable to GPT-6 Luna (max, 38) and DeepSeek V4.1 Flash (max, 39)2. On the firm's Cyber Index it scored 50, level with GLM-5.3-Flash and ahead of Kimi K3 and DeepSeek V4.1 Flash (max), but behind MiMo-V2.6-Pro at 56. Its best cyber result was 82% on CyberGym-E2E-AA, against 79% for MiMo-V2.6-Pro and 78% for GPT-6 Luna (max).

On GDP.pdf, a document and image reasoning test, it scored 19%, on par with MiMo-V2.6-Pro and behind Kimi K3 (22%). Artificial Analysis called that an 18-point gain over Mistral Large 3, partly driven by API changes that now allow 100 images per request, up from eight for earlier Mistral models.

Cost and availability

Mistral's announcement page does not list pricing. Artificial Analysis reported standard preview pricing of $1.36 per 1M input tokens and $4.18 per 1M output tokens, with $0.14 per 1M cached input tokens. At those rates it put the cost of one Intelligence Index task at $1.13.

A 50% launch discount for the first two weeks lowers that to $0.57, which Artificial Analysis said is still above GLM-5.3-Flash ($0.25) and DeepSeek V4.1 Flash (max, $0.27).

Artificial Analysis said the weights are planned for release at the end of October, and that once released the model would rank among the top three open-weights models on its Cyber Index. The Artificial Analysis article carries no publication date, and Mistral's post is dated 5 November, so the stated timing conflicts with the announcement date. The open-weights version has not been tested, and the results above apply only to the preview.

NVIDIA says full-stack codesign raises AI inference output per HBM byte

NVIDIA's Thursday post frames AI memory as an efficiency problem across the AI factory's hierarchy and claims up to 30× higher tokens per dollar on Vera Rubin NVL72 versus GB300 NVL72, attributing gains to full-stack codesign rather than raw HBM capacity increases.

Mira Sethi · about 1 hour ago

Researchers post arXiv papers on AI overthinking, ARC-AGI-3 and biology

The first paper, "Routes of Reasoning: Reading and Reshaping Test-Time Compute in MoE Models", introduces a method to optimize test-time compute in Mixture-of-Experts (MoE) models by selectively editing expert router scores.

Mira Sethi · about 11 hours ago

OpenAI says model patched 948 Linux kernel flaws

The company attached unusually specific numbers to a security automation claim, but only in a social post. Missing details include the model name, kernel scope and any outside review of the reported patches.

Julian Park · about 18 hours ago