Clarity for AI buildersSponsored by AI Routing
availability

Claude Fable 5.2 leads reported AI benchmarks, with higher peak task costs and restricted Mythos access

Claude Fable 5.2 leads reported AI benchmarks, with higher peak task costs and restricted Mythos access
Anthropic’s Claude Fable 5.2 leads Artificial Analysis’s reported AI index, according to The Decoder. The report describes Fable 5.2 and Mythos 5.2 as sharing a model foundation but differing in access and safeguards. Fable is available to all users, with restrictions on risky cybersecurity and biology tasks. Mythos removes certain restrictions and requires organizations to apply through Anthropic’s Verification Programs for Cybersecurity or Life Sciences. Anthropic continues to recommend Opus 5.5 for everyday use rather than positioning the new models as its default choice. According to Artificial Analysis results cited by The Decoder, Fable 5.2 scores 65 on Intelligence Index v5 at its highest compute tier. That places it ahead of Gemini 4 Carbon at 60, Opus 5.5 at 59 and Fable 5.1 at 50. The report notes that Gemini 4 Carbon had not yet been released. Fable 5.2 leads five of the index’s nine individual tests, including scores of 80 percent on HLE-Diamond and 73 percent on Terminal-Bench-Science. That overall lead does not extend to every workflow. On AA-Briefcase, which evaluates multi-step knowledge work, Fable scores 1,785 Elo, behind Sonnet 5.5’s 1,813 and within the confidence interval of Opus 5.5’s 1,800. On Terminal-Bench 4.0, Fable ties Opus at 60 percent, below Kimi K3.1 Pro, Gemini 4 Carbon and Sonnet 5.5. These results make the relevant task benchmark more informative than the aggregate ranking alone when choosing a model for a particular workflow. Listed pricing is $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the predecessors. Yet Artificial Analysis reports average per-task costs of $26.50 at Fable 5.2’s highest tier, compared with $10 for Opus 5.5 and $12.99 for Fable 5.1. Lower token prices therefore do not establish lower task costs: the amount of computation and token consumption matters as well. The lower compute settings offer a different trade-off. The Decoder reports that Fable 5.2 scores 62 at both High and Xhigh, still above every other model’s reported index result at any tier. It also cites 98,000 output tokens per task at these settings, compared with 173,000 for Opus 5.5 at Max. Together, the results suggest that evaluating effort settings is part of assessing the upgrade, rather than treating maximum compute as the only meaningful configuration. For cybersecurity, Anthropic says Mythos 5.2 improves vulnerability discovery, reproduction and remediation. Its internal CVE Rediscovery test uses project snapshots from before known vulnerabilities were fixed. In an October 2026 run covering 1,829 CVEs, Anthropic reported that Mythos 5.2 found 10.2 percent more vulnerabilities than Opus 5.5 and 14.4 percent more than Mythos 5.1. Those security claims should not be read as capabilities available without restriction in Fable. The report says the public variant blocks most security-related tasks, including penetration testing and exploit-code generation. Access policy is consequently a practical part of the distinction between the two models, not merely a difference in benchmark scores. Anthropic also reports a mathematical demonstration in which about 1,000 Fable 5.2 agents formalized the proof of Fermat’s Last Theorem in Lean in 30 hours. In the company’s comparison, 1,000 Opus 5.5 agents completed roughly 26 percent after 18 hours and made little further progress through hour 30. This is a reported formalization of an existing proof, not a claim that the model discovered the theorem or its original proof. The supplied evidence is The Decoder’s report, which combines cited third-party benchmark results with Anthropic’s own tests and demonstrations. The security and mathematics findings remain company-attributed claims here. The strongest supported practical conclusion is narrower than an across-the-board performance lead: Fable improves the reported aggregate score, while task-specific results, compute settings, costs and access restrictions still determine its suitability.