AI Agents Pass Just 1 in 4 Real Work Tasks, Berkeley Study Finds

OpenAI's Codex tool running on its GPT-5.5 model posted the best result, completing 26.2% of all assignments correctly -- roughly one in four. On the hardest tier of tasks, the kind requiring an AI to track many steps over a long stretch and get every piece right, pass rates across all tested systems averaged just 2.6%. Codex managed 8.6% on that tier. Anthropic's Claude Code tool, running on its Opus 4.7 model, failed every one of those hard assignments.

One version of the test threw enormous resources at the hardest problems: $630 worth of computing power and 763 million tokens, the units AI systems use to process text. It succeeded just 2.9% of the time. The researchers said the result shows that spending more money and computing power does not reliably make an AI agent better at this kind of work.