ProgramBench: Top LLMs Struggle to Rebuild Codebases, GPT 5.6 Solves Only 1%

jyangballin · x · 2026-08-11

Meta Superintelligence Labs, Stanford, and Harvard introduced ProgramBench, a benchmark evaluating whether language models can construct codebases from scratch. Given only a compiled binary and documentation, agents must architect and implement a complete codebase reproducing the original program's behavior across 200 tasks.

Leaderboard data reveals that current frontier models struggle significantly: the top-performing GPT 5.6 achieves a mere 1.0% full resolution rate, while Claude Opus 4.8 scores 0% (but passes 95% of tests in 16.5% of instances). Researchers noted that models construct genuinely novel solutions rather than just regurgitating source code.

Related event: GPT 5.6 Sol Tops ProgramBench, Halving Costs but Showing Python Bias(6 posts)→

Original post →

More from coding & agent

coding & agent channel →