TasteVal: AI Surpasses Human Experts at Designing Experiments

P-Zero Research's TasteVal benchmark found Claude Opus 5.5 reaches expert-level research judgment using about 43% of the compute and roughly 1/30 the per-attempt cost of human experts. Its compute multiplier of 2.30 was measured against 24 researchers from OpenAI, Google DeepMind, Nvidia and Microsoft Research.

• Claude Opus 5.5 scored a 2.30 compute multiplier, with a 95% CI of 1.15–4.37
• Research-taste capability has doubled every ~3 months since December 2025, versus ~14 months before
• Each run is capped at 40 GPU-hours on a single Nvidia H100
• The benchmark excludes problem selection, measuring only experiment design and interpretation