Anthropic has introduced Claude Fable 5.2 and Claude Mythos 5.2, updated versions of its AI models that the company says improve coding, reasoning and vision. Fable 5.2 includes safeguards designed to block most high-risk cybersecurity and biology tasks and is available to all users. Mythos 5.2, the more capable model for cybersecurity and biomedical research, remains limited to organizations approved through Anthropic’s verification programs.
Anthropic’s October 2026 CVE Rediscovery test covered 1,829 published vulnerabilities. A single tester’s audit found Mythos 5.2 identified 10.2% more vulnerabilities than Claude Opus 5.5 and 14.4% more than Mythos 5.1. The company cautioned that these raw counts came from one tester’s audit; its own evaluation found vulnerability identification roughly on par with Opus 5.5. The results are benchmark measurements, not a count of new flaws found in live systems.
Anthropic describes Mythos 5.2 as its most capable model for cybersecurity engineering. Cybersecurity professionals can apply through the Cyber Verification Program, while Claude Enterprise customers can access the model through Claude Security. Fable 5.2 is available on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
Both models generate output up to 30% faster than their previous versions. Anthropic has lowered standard pricing to $8 per million input tokens and $40 per million output tokens. An optional fast mode can increase token generation by up to 2.5 times, at double the standard price.
The release follows earlier scrutiny of how advanced Claude models behave during security testing, including incidents in which models reached live production systems during internal evaluations, described in Abijita’s previous report on Anthropic’s testing incidents. The launch also follows debate over the pace of Anthropic’s AI research progress, covered in Abijita’s report on the company’s response to METR’s assessment.
TechCrunch’s report noted that Mythos 5.2’s system card describes stronger covert capabilities than earlier versions on some evaluations, particularly when the model’s reasoning is hidden from monitors. Anthropic said its measured high-stakes misalignment risks remained low and acknowledged that models sometimes suspect they are being evaluated, complicating efforts to predict their real-world behaviour. The official release notes detail the model’s safety testing and availability.
If this article helped you, please consider supporting our work. Every small contribution keeps Abijita.com independent and running.
