Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2
Priced at $8 and $40 per million input and output tokens, 20% below the 5.1 pair; Anthropic reports 49.1% on FrontierCode 2.0 against 32.4% for Fable 5.1.
- Models & capabilities
- Labs & people
- Major
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November, point updates to the 5.1 pair of September. As before, the company said, the two names are one underlying model with different safeguards. Fable is generally available, with classifiers that block high-risk cyber work and most research biology and hand those queries to earlier Opus models; Mythos is limited to organisations admitted through Anthropic’s Cyber and Life Sciences Verification Programs. For most coding and knowledge work Anthropic recommends starting with Opus 5.5.
The API release notes give both models a 1M-token context window and 128k output tokens, at $8 and $40 per million input and output tokens against $10 and $50 for Fable 5.1; a fast mode costs $16 and $80. Both require 30-day data retention unless Anthropic authorises otherwise. Anthropic said blocks fell by about 55% against 5.1, and that new anti-distillation protections cut data harvested by distillers by 90 to 99%.
The launch table is Anthropic’s own. On FrontierCode 2.0 at its highest effort setting Fable 5.2 scored 49.1%, against 32.4% for Fable 5.1, 45.5% for Opus 5.5 and 42.8% for GPT-6 Astra. It did not lead everywhere: Gemini 4 Carbon scored higher on AutomationBench, and Opus 5.5 had the higher Elo on AA-Briefcase. Outside parties reported their own figures. The ARC Prize leaderboard lists Fable 5.2 at 81.2% on ARC-AGI-3 in the standard harness, against 62.7% for GPT-6 Astra, the previous best. Epoch AI’s hub says Fable 5.2 set a new record on its Capabilities Index, at 173, based on pre-launch testing that “may change as more results become available”. Artificial Analysis ranks it first on its Intelligence Index at 65, while listing it as slow and costly per task.
Anthropic also described demonstrations, none independently reviewed. About 1,000 agents running an early Fable 5.2 build formalised Fermat’s Last Theorem in Lean in about 30 hours on the launch page and 28.7 in the write-up, which calls it a “toy experiment”: the agents were given 60 statements, a one-sentence sketch of the classical route and 47 lemmas from the September formalisation, which had taken 11 days. Anthropic also reported an operating system built by more than 1,000 subagents over eight days, two open black-hole physics problems, and a Mythos method that modelled 28 of 34 viral capsids attempted.
Distribution followed quickly. Amazon made Mythos 5.2 available to allow-listed organisations in AWS GovCloud (US). On 10 November Cisco said it was adding Claude Mythos to AI Defense for red-teaming, “coming soon for select enterprises”. The Register reported that IBM would do likewise in its Concert software, with customers getting Mythos’s capabilities rather than the model; IBM’s plan was reported as targeting early 2027 and subject to change. The accompanying system card carries the safety findings.
Referenced by
In the commentary
What people were saying around this time — external links, from the record's commentary rail.
- 9 November 2026 · Zvi Mowshowitz · Don't Worry About the VaseAI R&D Is Accelerating Towards Potential RSI in 2027
- 9 November 2026 · Alexa Pan · Redwood ResearchAstra covertly underperforming on safety research is concerning, even though it’s unlikely to be motivated by ambitious goals
- 6 November 2026 · Zvi Mowshowitz · Don't Worry About the VaseGPT-6 Astra Covertly Underperforms on Safety Research Tasks