Epoch AI's MirrorCode benchmark asks a model to rebuild a whole program from scratch, without seeing its source code, until its output matches the original exactly on visible and held-out tests. The leaderboard setup is 30 tasks (15 programs, each in two languages), each run 3 times. Leaderboard: https://epoch.ai/benchmarks/mirrorcode
State on 2026-11-04 (Epoch's downloadable data, mirrorcode.csv in https://epoch.ai/data/benchmark_data.zip):
Top score: Claude Opus 5.5 (high) at 78.9%
Next: Claude Fable 5.1 (high) 73.3%, Claude Fable 5 (high) 63.9%, Gemini 4 Argon (high) 58.9%, GPT-6 Astra (high) 46.7%
Question: what is the highest score any model has on this leaderboard at close?
Resolution: On or shortly after Dec 1, 2026, I'll take the maximum score shown for any model in mirrorcode.csv / on the Epoch MirrorCode leaderboard. Only runs whose evaluation started on or before Nov 30, 2026 (UTC, the CSV's "Started at" column) count. The score is the percentage Epoch shows (the CSV's mean score × 100). If the page and the CSV disagree, the CSV wins. A score exactly on a bucket edge goes in the higher bucket (80.0% counts as "80% to under 85%"). If Epoch renames the benchmark, I'll use its direct successor. If they change the task set, I'll go by the score on the current 30-task setup where that can be told apart; if it can't, I'll explain in the comments before resolving.
Note: with 90 graded attempts, one more success on top of today's leader (72 of 90) is exactly 80.0%.
1,000My guess: ~50% it stays under 80%, 18% for 80–85, 17% for 85–90, 15% for 90+. The upside is a new Claude Fable: Anthropic holds the top 3 here, Manifold has a new one by Nov 21 at ~77% (https://manifold.markets/Terminator2/when-will-anthropic-release-its-nex), and Epoch tested Fable 5.1 nine days after launch. A release date or a non-Anthropic model above 75% would move me.