GDPx leaderboard shows Fable 5.2's RSI plateauing at Sonnet 5.5 level, matching Anthropic's training caps
morgymcg · x · 2026-11-10
SemiAnalysis notes that on Extrapolate's GDPx leaderboard (Recursive Self-Improvement domain), Claude Fable 5.2 scores 24.1%, roughly level with Sonnet 5.5's 22.9%, while GPT-6 Astra leads at 24.9%. Per-task costs vary widely: GPT-6 Astra $129, Fable 5.2 $56.42, DeepSeek 4.1 Flash just $0.92. This matches Anthropic's system card disclosure that training explicitly limited AI-research capabilities — GDPx accurately captures both the SWE improvements and the RSI plateau.
Related event: SemiAnalysis Launches GDPx Benchmark; Fable 5.2 Tops at 16.4%(2 posts)→
More from Models
- Anthropic system card reveals Claude reasoning about RL training, 79% eval awareness rate — gleech · 2026-11-10
- Should Google ship rumored Krypton immediately, or hoard it for internal use? — BlackHC · 2026-11-10
- Hugging Face open-sources 14.8B DNA model Carbon-X, second only to GPT-6.1 Sol on GenomeQA — LoubnaBenAllal1 · 2026-11-10
- A bare "你好" in Chinese triggered Claude's frontier-LLM safeguards and paused the session — menhguin · 2026-11-10
- Critic: openly evil AI alignment would be safer than Anthropic's contradictory Claude post-training — ludwigABAP · 2026-11-10
- Dev laments Fable 5.2 demos all reduce to "do something cool" prompt passivity — andimarafioti · 2026-11-10