Claude Fable 5.2 takes #1 LLM spot at 65 on Artificial Analysis

The launch arrives the same day Anthropic's $ANT IPO roadshow began, a timing that several commentators read as deliberate. SemiAnalysis questioned whether a stronger release will come before the listing, while others doubted the model is good enough to save the IPO. Within the release itself, Anthropic has imposed deliberate limits on RSI, or recursive self-improvement, capabilities. The system card states that Anthropic's most capable model remains internal, implying the publicly released Fable 5.2 sits below what the lab has built.

Independent evaluations sketch a model that is broad but not uniformly dominant. Fable 5.2 set new records on HLE-Diamond at 80%, AA-Omniscience at 67, Terminal-Bench-Science at 73%, CritPt v2 hard at 77%, and FrontierCode 2.0 at 49%. Yet it trailed Sonnet 5.5 on AA-Briefcase Elo and GDP.pdf, and sat behind Gemini 4 Carbon and Kimi K3.1 Pro on Terminal-Bench 4.0. The cost picture complicates the 20% per-token price cut: at max effort, cost per task is about $26.5, roughly 2.6× Opus 5.5, because Fable 5.2's testing breadth far outstrips earlier models, generating hundreds to thousands of test cases where Opus 5.5 and Fable 5.1 stopped at about 20.

Subscription value reshapes how existing users will experience the release. One analysis of 1.1M logged requests on a $200 Ultra subscription found the plan yields about $15,000/month of Opus 5.5 but only about $3,000 of Fable 5.2, and Fable is capped at 50% of a weekly limit — so Fable-only users effectively lose half their limit. Anthropic's own staff offered mixed guidance: Felix Rieseberg said Fable beats Opus on big, long tasks such as whole-codebase rewrites, but still often prefers Opus 5.5 day-to-day because Fable thinks a little slower. Community testing surfaced the same split, with some practitioners calling it a jagged AGI while others found it making stupid mistakes and not significantly better than Opus 5.5 for the price.