Claude Fable 5 Outscores GPT-5.6 in From-Scratch Code Rewriting. Here's Why

Epoch AI updated its MirrorCode leaderboard, revealing that Anthropic's Claude Fable 5 solved 64% of tasks, while OpenAI's GPT-5.6 Sol managed only 20%. This threefold difference looks anomalous given the models perform comparably on other tests, but Epoch AI and Anthropic offered no official explanation. MirrorCode, developed with METR, requires models to rewrite entire programs from scratch, seeing only an executable 'black box', documentation, and visible tests; solutions must pass all hidden tests. The benchmark includes 15 real programs like sed and Ruff, each implemented in two languages, with a budget of 10 billion tokens and 7 days per attempt. Analysis of the heatmap shows Sol often partially solves tasks (50%, 33%, 17% success rates), while Fable nearly always completes them (100%, 83%). Another clue: Fable's performance drops only slight on Ada (61% vs 64% on Go), while OpenAI models lose significantly on Ada — Sol falls from 24% to 19%, GPT-5.4 from 21% to 12%, and GPT-5.5 from 17% to 5%. Since almost no Ada implementations of these programs exist in training data, Fable's Ada resilience indicates genuine skill rather than memorization. However, Epoch's own contamination check found signs of memorization in 17 of 25 programs, and the benchmark's exact-specification setting differs from real-world development.