Claude Fable 5.2 Official ReleaseProvider report
2026-11-09From the complete main comparison table on Anthropic's November 9, 2026 Claude Fable 5.2 / Mythos 5.2 release page, supplemented with results the release table omits, taken from System Card section 8.1 and body text (sourced via raw.system_card_section). The Terminal-Bench-Science 0.1 Fable / Mythos dual value and the AA-Briefcase v1.1 rubric / Elo dual value are split, recording 25 metrics and 73 valid scores in total; em-dash cells generate no scores. Claude Mythos 5.2 and Fable 5.2 are the same model, so no separate model is created. CyScenarioBench, CVE Rediscovery and De Novo Binder Design report only Mythos-mode or no-safeguard results; under the consolidation rule they are attributed to Claude Fable 5.2 / 5.1 with the evaluation mode recorded in raw. CyScenarioBench is Anthropic's run of Irregular's 10-challenge subset; because each lab scores on unique tasks, results are not comparable across labs, and GPT-6 Cyber's 61% is not imported since it is neither a covered model nor the same task set. The Gemini 4 Carbon column is omitted because no profile for that model exists on the site yet. Unless otherwise noted, scores use each model's highest available effort; FrontierCode 2.0 shows the best score across effort levels (xhigh for Fable 5.2 / 5.1). OSWorld 2.1 uses the official offline set (82 tasks) in both partial and strict forms, and AutomationBench uses Zapier's private held-out set v1.0.6; all are stored as separate criteria and never overwrite existing OSWorld 2.1 or legacy variants. System Card comparison-model figures identical to scores already on the site are not re-imported; fresh Elo rounds such as GDPval-AA v2.1 coexist with older snapshots.