Anthropic Says Claude Fable 5.2 Beats GPT-6 in Cyber Tests

Nov 10, 2026: Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, claiming improvements in coding, reasoning and vision. In Irregular’s CyScenarioBench tests, Fable 5.2 completed 76.3% of multi-stage cyber operation challenges, versus 61.7% for Fable 5.1 and 61% for OpenAI’s GPT-6 Cyber. On the same benchmarks, Google's Gemini 4 Argon solved 12 challenges end to end in a related report, up from 5 for Gemini 3.8 Flash Cyber.

The article notes that on CyScenarioBench, GPT-6 Cyber completed 11 of 11 long-horizon scenarios at least once, with an average success rate of 61% (up 7 percentage points), and on Atomic Challenges, GPT-6 Cyber solved all 22 challenges, achieving 97% in vulnerability research and exploitation, 98% in network attack simulation, and 63% in evasion.

Anthropic also reported Claude Fable 5.2 scored 89.6% on the CVE Rediscovery benchmark, its internal assessment of whether an AI model can independently rediscover known vulnerabilities from affected source code, drawn from 1,829 advisories for open‑source projects published in October 2026. The company described Claude Mythos 5.2 as its most capable model for cybersecurity engineering, capable of flagging more vulnerabilities while improving reproduction and fixing them.

The models’ cyber capabilities are accessible through the Cyber Verification Program and Claude Fable 5.2 is available on AWS, Google Cloud and Microsoft Azure. The piece cites Anthropic’s system card and pre-release training results claiming assistance in complex software migrations, device prototyping, black hole physics problems and Fermat’s Last Theorem formalisation within 30 hours.

The article also notes Anthropic updated Claude Security with continuous code scanning and automatic fix tracking to help security teams monitor changes and prioritise live fixes, with findings accompanied by severity ratings and explanations of exploiable paths.

It is stated that the 5.2 release continues to address alignment and misuse resistance, with Fable continuing to show lower likelihood of publishing harmful content to public package registries when not prompted, though incidences occurred under simulated user encouragement. The report presents these claims as from Anthropic, Irregular and the Cyber Verification Program, and preserves the specific performance metrics and CVE-related benchmarking figures as stated by the sources.