The new TasteVal benchmark evaluates how effectively AI models design experiments and interpret results, a capability P-Zero Research defines as research taste. The test covers eight tasks, including fine-tuning, training data selection, and defense against adversarial prompts, using a fixed coding agent on a single Nvidia H100 GPU.
Data from P-Zero Research indicates that the research taste of leading AI models has doubled approximately every three months since December 2025. The top-performing model, Anthropic's Opus 5.5, achieved a compute multiplier of 2.3, matching human expert performance with less than half the compute resources.
The human baseline for the benchmark was established by 24 experts from organizations including OpenAI, Google DeepMind, Nvidia, and Microsoft Research. Each experiment run was limited to 40 GPU-hours or 120 hours of wall-clock time, with costs per run averaging one-thirtieth of a human expert attempt.
Jack Clark, co-founder of Anthropic, has adjusted his forecast for recursive self-improvement in AI from 2028 to 2027. This shift follows observations of progress in mathematical breakthroughs and performance on the TasteVal benchmark, despite the test's narrow focus on specific research tasks.
While AI agents have shown the ability to contribute up to 55 percent of methodology proposals in other studies, humans currently retain control over more than 85 percent of final research decisions. The industry remains divided on the timeline for fully autonomous AI research capabilities.
