Claude Mythos 5.2 review, pricing and limits

Anthropic's launch table labels three scores as Mythos 5.2 results, all vendor-reported and all at the highest available effort. On CVE Rediscovery, an internal test of how many published vulnerabilities a model finds from a snapshot taken before the fix, it scores 89.6% on an October 2026 run against 1,829 CVEs, against 78.3% for Mythos 5.1 and 81.3% for Opus 5.5 without safeguards. On CyScenarioBench it scores 76.3%, against 61.7% for Mythos 5.1 and 67.6% for Opus 5.5 without safeguards; Anthropic notes that each lab's CyScenarioBench scores rest on different tasks, so the 61% it lists for GPT-6 Cyber is not a like-for-like comparison. On Terminal-Bench-Science 0.1 it scores 75.3%, against 74.6% for Fable 5.2 and 58.7% for Opus 5.5. No independent leaderboard lists Mythos 5.2 because access is gated, so Hokai stores none of these scores in its benchmark fields.

The testimonials in the same post are Anthropic's selection. Mozilla's Brian Grinstead says Mythos 5.2 roughly doubled the vulnerability discovery metrics of Mythos 5.1 across their respective scanning campaigns, and Rubrik's chief technology officer says the company would switch its Code Guardian system from Mythos 5.1 to 5.2. The same post, dated 9 November 2026, says an unnamed security company found that one in six confirmed bugs from Mythos 5.2 had been missed by every other frontier model it tested; that is a vendor-reported claim nobody outside Anthropic has checked. Anthropic's own footnote adds a limit: those counts come from one tester's audit, and on its own evaluations vulnerability identification is roughly on par with Opus 5.5.

It reads text and images and writes text, with a 1 million token context window and up to 128,000 output tokens. Adaptive thinking is always on, effort runs from low to max with high as the API default, and the knowledge cutoff is June 2026. Computer use goes through the computer_toolset_20260801 toolset. Anthropic's launch post of 9 November 2026 says output is generated up to 30% faster than on Mythos 5.1 (a vendor figure with no independent measurement), and an optional fast mode raises that to up to 2.5 times, but fast mode for Mythos is open only to organizations in the verification programs.

Access is by verification. Cybersecurity teams apply through the Cyber Verification Program, whose access comes in three tiers named Defense, Red Team and Specialized, while life-science researchers go through a separate Life Sciences Verification Program. Claude Enterprise customers can also use Claude Security, which scans codebases with Anthropic's most capable models including Mythos 5.2, without joining the program. Verified organizations call the model on the Claude API, Claude in Amazon Bedrock, Google Cloud (Anthropic-managed and Google-managed) and Microsoft Foundry, and Anthropic lists Claude Code among the places fast mode is offered.

Anthropic says the risk of unsafe behaviour in high-stakes situations is low, as with its other recent models. On an automated behavioural audit of about 1,900 scenarios, Mythos 5.2 scored 2.67 overall, where lower is better, against 3.28 for Mythos 5.1, 2.79 for Opus 5.5 and 3.10 for Claude Sonnet 5.5. In two earlier evaluations it also made substantially fewer sandbox-escape attempts than any other model Anthropic has tested. The same post concedes a problem: Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any recent Claude model, which makes it harder to know how they behave in real deployments. The system card has the full results.

Most teams should not start here. Fable 5.2 is the same model without an application, and Anthropic's own guidance is to begin on Opus 5.5 for most coding and knowledge work. Teams that want a rival cyber-focused model have GPT-6 Cyber, which Hokai lists as gated behind its own vetting programme. Anything that needs zero data retention is a poor fit by default, because Mythos 5.2 is a Covered Model with a 30-day retention requirement.

Anthropic gives a training data cutoff of June 2026 and does not publish the architecture or parameter count. The docs call for 30 days of retention so safety monitoring can run, and zero data retention is off the table unless Anthropic expressly authorizes it, though its Enterprise Frontier Safeguards beta, which runs the monitoring on data stored in the customer's own infrastructure, is built to work with zero retention. Anthropic's commercial terms bar it from training models on what customers send through its services.

Against Mythos 5.1 the changes are a lower token price, faster output, an optional fast mode and four breaking API changes: editing anything before a thinking block now returns a 400 error, only Fable 5.2 and Mythos 5.2 can read each other's thinking blocks, the older computer_20251124 tool fails on every platform except Amazon Bedrock, and an advisor paired with a Mythos 5.2 executor must itself be Fable 5.2 or Mythos 5.2. Anthropic says the model stays available until at least 9 November 2027.