3x-repeat benchmark: Fable 5.2 tops Opus 5.5 on most protein contact prediction tasks
sokrypton · x · 2026-11-19
Follow-up from sokrypton: each task repeated 3x across 3 Claude-family models, comparing top-L precision on protein contact prediction:
- Fable 5.2 improved over Fable 5.1 and beats Opus 5.5 on most proteins: ubiquitin 0.85 vs 0.66, XisI 0.68 vs 0.32, thioesterase 0.34 vs 0.19.
- Opus 5.5 only leads on L29-like (0.29) and 6MRR (0.52).
- Run-to-run variance remains high (Fable 5.2's Top7 runs: 0.53/0.79/0.83), confirming the earlier non-reproducibility observation.
More from Models
- GAO audit finds Claude, Grok, ChatGPT and Gemini safeguards easily bypassed — midwoodgirl10 · 2026-11-19
- Harmonic's Aristotle rewrites Bitcoin Core's C signature parser in Rust with a formal proof — satnam6502 · 2026-11-19
- GPT-6.2 Astra users note it can work 30 minutes without any progress updates — Angaisb_ · 2026-11-19
- Arena launches six new agent indexes from millions of real sessions, updates score scaling — arena · 2026-11-19
- Arena points readers to the live Agent Arena Index and its methodology — arena · 2026-11-19
- Investor complains of 'rude' Anthropic rate limit, blames Nvidia deal monopoly — StewartalsopIII · 2026-11-19