Claude Opus 5.5 fully rebuilt 37 of 200 real programs on ProgramBench — 18.5%, up from a best score of zero when the benchmark launched in May — working from only the finished program and its manual, with no source code, no internet, and no peeking inside. The next-best model reached just 9%, and 70.5% of tasks were only nearly solved, so human checks still decide what can be trusted.
Hook: Claude rebuilt programs from scratch — 18.5%, up from zero.
Context: ProgramBench hands AI a finished program and its manual only — no source code, no internet, no peeking inside.
Stakes: Claude Opus 5.5 fully rebuilt 37 of 200 programs, double the next model at 9%; at launch in May the best score was zero.
Payoff: Zero to 18.5% — the rebuild era just started. Send this to one person who writes code.
Watch next: Google’s Genie 4 Turns Text Into Playable Worlds — https://www.instagram.com/p/DfA0einAzBV/
Instagram reel: https://www.instagram.com/reel/DfBaJPrguRB/
Sources: ProgramBench leaderboard — https://programbench.com/ ; Import AI 478 (Jack Clark, Oct 26, 2026) — https://jack-clark.net/2026/10/26/import-ai-478-math-for-alignment-programbench-and-mirendil-shows-what-rsi-means/
Leave a Reply