Jeremy Berman Claims 96.2% on ARC-AGI-3 with Opus 5
Setup uses Claude Code, one action command, and filesystem logs with almost no ARC-specific tools.
TLDR
Jeremy Berman posted that he scored 96.2 percent accuracy on ARC-AGI-3 and 99.3 percent pass@2. He said the program consists of Claude Code plus Opus 5 at high setting, one action command, and filesystem logs. The approach required almost nothing ARC-specific. Berman linked a GitHub repository containing the code. Several engineers and researchers replied to the post. Some described the result as evidence that general coding harnesses suffice for the benchmark. Others noted the method still falls short of general intelligence.
Combined views
474.3K
18 Sources, first seen 49d ago