Mistral Releases Large 4 Preview, Claiming a Cybersecurity Edge Over Chinese Rivals

Mistral Large 4 trails Chinese open models in independent tests, despite cybersecurity strength

Mistral’s newest AI model has a cybersecurity standout, but not an overall lead over Chinese open models. Artificial Analysis gave Large 4 Preview a score of 38 on its Intelligence Index, below five Chinese rivals. Its strongest security-test result sits alongside a less flattering comparison: higher benchmark task costs than several competing models.

The evaluation gives substance to the competitive claims surrounding Mistral’s October 6 release. Artificial Analysis calls Large 4 the highest-scoring model from outside the United States and China on its index. That distinction describes a geographic lead within this evaluation—not leadership across all models or every kind of work.

The Intelligence Index combines 10 evaluations, including professional document reasoning, coding, knowledge and long-context reasoning. Large 4’s score matches GPT-6 Luna at its maximum setting. The Chinese open-weight models ahead of it span a range of scores, rather than one narrowly superior competitor:

Xiaomi MiMo-V2.6-Pro: 46.

Z.ai GLM-5.3 at maximum effort: 45.

Moonshot Kimi K3 at maximum effort: 44.

Z.ai GLM-5.3-Flash: 42.

DeepSeek V4.1 Flash at maximum effort: 39.

There is also an availability distinction. Those competitors have open weights; Large 4 does not yet. Mistral offers a public API preview and says it will release the weights by the end of October 2026. Artificial Analysis currently labels the preview proprietary.

Cybersecurity provides the clearest counterpoint to the overall ranking. Large 4 scored 50 on Artificial Analysis’s Cyber Index, tying GLM-5.3-Flash but trailing MiMo-V2.6-Pro’s 56. Its strongest component was CyberGym-E2E-AA, where it finished ahead of both MiMo-V2.6-Pro and GPT-6 Luna.

Mistral describes the CyberGym task as reproducing a real vulnerability in open-source software and then patching it. That makes the result relevant to a particular security workflow. But it is only one of three tests in the Cyber Index, and Large 4’s results on the other two are substantially lower.

Refusal behavior also complicates the comparison. Artificial Analysis’s chart marks Claude Opus 5.5 and GPT-6 Astra as declining some tasks on safety grounds. Mistral argues that such refusals can obstruct legitimate vulnerability research and incident response. Their near-zero CyberGym scores therefore reflect whether they would undertake the work, not simply whether they could solve it.

The company is testing the model with cybersecurity leaders, vetted partners and state authorities before releasing weights. Mistral says those participants receive reduced moderation and expanded cyber capabilities. That controlled testing is separate from the ordinary public preview.

Large 4’s standard API prices are $1.36 per million input tokens and $4.18 per million output tokens. Tokens are the units used to meter model input and output. Artificial Analysis lists those rates below its comparison medians, yet measures an average cost of $1.13 per Intelligence Index task.

The distinction is usage. Large 4 generated 200 million output tokens during the evaluation, against a comparison median of 81 million. Artificial Analysis calculates task cost from token charges, divides by task count and weights each evaluation’s contribution. Its figure is a benchmark-specific measure, not a promised bill for a customer’s workload.

A 50% discount for the first two weeks cuts Large 4’s benchmark task cost to $0.57. Even then, it remains above GLM-5.3-Flash at $0.25 and DeepSeek V4.1 Flash at $0.27. MiMo-V2.6-Pro costs $0.13 per task in the same comparison while posting the highest Intelligence Index score among these models.

Large 4 does offer faster-than-average generation: Artificial Analysis measured 116.1 output tokens per second, compared with 86 for its comparison average. It accepts text and images and returns text. The planned weights release would also allow private-cloud or on-premises deployment, giving organizations a different access model from today’s hosted preview.

Mistral Begins Reduced-Moderation Cyber Testing Before Large 4 Weights Release

Mistral is testing Large 4 with cybersecurity leaders, vetted partners and state authorities before releasing its downloadable weights, giving those testers reduced moderation and expanded cyber capabilities. In its October 6 announcement, the company makes two distinct claims: the model can reproduce real software vulnerabilities, yet refuses malicious cybersecurity requests more often than open-source models in its safety comparisons.

The most concrete cybersecurity result concerns a task that combines reproducing a real vulnerability in open-source software with patching it. Mistral reports an 82% score on that component of the Artificial Analysis Cyber Index, calling it the highest score of any model. That is one test, not an overall cybersecurity ranking.

Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because they refuse the task. Its argument is that defenders sometimes need to demonstrate a vulnerability before repairing it. A refusal can therefore prevent legitimate security work, rather than indicate an inability to solve the problem.

The company also reports that Large 4 solves 93% of Cybench’s 40 security-competition exercises. In internal testing, Mistral says it proved useful for analyzing malware, prioritizing vulnerabilities and writing detection rules. Those examples extend its security pitch beyond the vulnerability-reproduction benchmark, but remain company-reported results.

Mistral separately measures whether Large 4 rejects malicious requests. It says the model’s average refusal rate on cybersecurity prompts from JailbreakBench, StrongREJECT and AgentHarm is higher than all open-source models in its comparison. This is distinct from refusing the vulnerability-reproduction task: the safety claim concerns requests classified as malicious.

Another safety measure addresses indirect prompt injection—attempts to steer an AI through instructions it encounters in outside material. Mistral reports 93.3% attack resistance on Lakera’s B3 benchmark. The partner testing uses the same model with reduced moderation, so its access conditions should not be confused with the model’s general safety results.

For security teams, Mistral’s deployment argument is about who controls access and refusals. It says provider-level restrictions can interrupt legitimate research or incident response. Planned open weights and self-deployment would let organizations run security work under their own policies, with private-cloud and on-premises use also planned.