Mistral said it launched a public preview of Mistral Large 4 and Anthropic said it expanded its Cyber Verification Program to three access tiers, according to separate company announcements.

The updates matter because both companies are framing advanced model access around cybersecurity work: Mistral is tying its new model to open-weight deployment and sovereign infrastructure, while Anthropic is widening controlled access to models with reduced cyber blocking for vetted defenders.

Mistral said in its announcement that Mistral Large 4, or ML4, is a 1 trillion-parameter multimodal model with 52 billion active parameters, available now through a preview API in Mistral Studio, with weights due by the end of the month. The company said the model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral datacenters in Europe and that the preview is served on the same infrastructure.

Mistral's post presents a long list of benchmark and evaluation results, but those results are vendor-supplied. The announcement says ML4 scored 82% on one Artificial Analysis Cyber Index test, solved 93% of Cybench challenges, and scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. It also says the model scored 59.9% on AutomationBench and reached 1,393 Elo on AA-Briefcase.

The same announcement says Mistral is still red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities before releasing the weights. It also says the company will share further details on the architecture, additional benchmarks and post-training methodology later. The announcement includes no independent test results.

Anthropic adds tiers for reduced-blocking cyber access

Anthropic said its updated Cyber Verification Program, or CVP, combines its earlier Project Glasswing and CVP efforts into a single program with Defense Access, Red Team Access and Specialized Access tiers. The company said the program makes advanced cyber capabilities and reduced blocking classifiers available to qualifying security professionals.

According to Anthropic, each tier includes access to Claude Opus 5.5, Claude Sonnet 5.5, Claude Mythos 5.1 and future models. The company said Defense Access covers uses such as incident response, malware reverse engineering and vulnerability analysis, while Red Team Access adds authorized penetration testing and red-teaming. Specialized Access is reserved for a limited set of verified organizations authorized to test systems such as power grids, telecom networks and interbank transfer infrastructure.

Anthropic also published company-run evaluation results for the new tiering system. It said that on five attempts across 10 CyScenarioBench challenges per access tier, the generally available model was blocked on every task at the first prompt; in Defense Access, 46 of 50 trials were blocked at some point and four succeeded; and in Red Team Access, no blocks occurred and Claude Opus 5.5 completed 34 of 50 tasks, which Anthropic said was effectively equivalent to a 67.6% success rate with no safeguards applied. Those results come from Anthropic's own announcement rather than an independent test.

Access conditions and deployment details

Anthropic said data retention is required for organizations enrolled in CVP so it can monitor cyber misuse. The company said that once Enterprise Frontier Safeguards becomes available later this fall, eligible organizations will be able to store data in cloud infrastructure they control. Anthropic said CVP is available on the Claude Platform, Google Cloud Vertex AI and Microsoft Foundry, and on Amazon Bedrock only for customers eligible for Enterprise Frontier Safeguards.

Anthropic said that through Project Glasswing, partners uncovered at least 129,000 verified software vulnerabilities between April and July 2026, while Anthropic's own open-source scanning found an additional 5,500 verified software vulnerabilities between April and October 2026. The company said more than 33,000 of those verified vulnerabilities have so far been rated critical- or high-severity, while also saying that figure is likely an undercount based on survey data from only a subset of partners.

The two issuer announcements establish Mistral's public preview for Large 4 and Anthropic's expanded cyber access program.