report · 2026-11-09
Claude Fable 5.2 and Mythos 5.2: AI research and development evaluation context
On this page
Why it matters here
Anthropic's November 9 system card updates its account of AI research and development capability for Claude Fable 5.2 and Mythos 5.2. It reports a revised capability-index fit that favors sustained faster growth over the previously preferred one-time jump, while retaining uncertainty about AI's causal contribution and judging that Mythos 5.2 does not fully substitute for its research staff. The card's quoted METR testing finds stronger verifiable-task performance on an earlier, more capable snapshot, with mixed evidence for open-ended research judgment. These are qualified evaluation and methodology updates, not demonstrated autonomous recursive improvement.
AI-R&D capabilityResearch autonomyEvaluation integrity
What to keep in mind
- Anthropic reports these evaluations and internal-use judgments; the tracker has not independently reproduced them or verified complete raw runs, costs, evaluator records or internal data. This entry covers the AI-R&D section, not the entire system card.
- The new AECI fit is dated November 6, 2026. It refits the scale and benchmark basket, so its values cannot be compared directly with older card fits or Epoch's public index. Its trend includes an unreleased internal model and an earlier Mythos 5.2 snapshot.
- A historical change in capability growth is different from AI-attributable progress relative to an unassisted counterfactual. Anthropic's uncertain policy-threshold judgment is not an independently established causal acceleration estimate or the tracker's risk assessment.
- The quoted METR pre-deployment testing concerns a snapshot stronger on AI R&D than the released model. It predates METR's separate, higher-access October assessment; the two reports and populations must remain distinct.
- Better performance on verifiable optimization tasks does not establish better open-ended research judgment, full researcher substitution or an improvement mechanism that reliably improves future improvement. The card retains early-stopping, verification and research-taste weaknesses.
- Anthropic did not run its saturated legacy task-based AI-R&D evaluations for this release and withholds CoBench results while revising that evaluation. Neither omission means zero performance or a new comparable score. No numeric observation or public figure is added here.