PostTrainBench: Fable 5.2 ties 5.1 at 44.6%, speculating Anthropic capped AI research ability
scaling01 · x · 2026-11-13
PostTrainBench added Fable 5.2 and Gemini 4 Carbon to its leaderboard: Fable 5.2 averages 44.6% (±1.2), exactly matching Fable 5.1 and trailing only the 48.4% Official Instruct model, while Gemini 4 Carbon scores 41.0% (±1.7), between GPT 6 (Astra) and Gemini 4 Argon (40.4%).
The poster speculates the flat Fable result stems from Anthropic deliberately training the model to underperform on AI research tasks, as disclosed in its system card. Elsewhere on the board: Opus 5.5 at 43.8%, GLM 5.3 at 34.5%, Kimi K3 at 31.5%, Grok 4.5 at 22.6%.
More from Models
- Perplexity's Decider V1.3 decision models free on OpenRouter until Nov 19 — antoine_chaffin · 2026-11-13
- Dev begs AI industry to bring back small/medium/large model naming — andpoul · 2026-11-13
- Microsoft Copilot adds Anthropic's Claude Fable 5.2 to Cowork and Copilot Studio — luisdans · 2026-11-13
- Gemini 4 sweeps TxBench antibody development leaderboard at 60%+, double Claude's score — kenbwork · 2026-11-13
- Users grow impatient as OpenAI keeps teasing a powerful unreleased model — imjustnewatai · 2026-11-13
- Pro user burns entire 5h limit on Fable in minutes without finishing the task — CutBulkMaintain · 2026-11-13