TasteVal - Performance Multiplier: leaderboard

The final-result companion to the TasteVal compute multiplier (P-Zero Research, 2026), on the same eight withheld AI R&D tasks and Researcher and coding-agent protocol. Each run scores the test result of its official submission after the full 40 GPU-hour budget, rescaled per task so the naive baseline is 0 and the best human expert attempt is 1, whatever compute it used; seeds are averaged and tasks combined geometrically. Higher is better, with no upper limit.

Metric: Performance multiplier against the expert baseline: test score of the official submission of each run within the 40 H100-hour budget, rescaled per task so the naive baseline is 0 and the best human expert attempt is 1 (runs below the naive baseline count 0), arithmetic mean over seeds and then geometric mean over the 8 withheld AI R&D tasks (a task with no run above the naive baseline enters at 0.01), regardless of compute used; same Researcher and fixed Coder protocol as the compute multiplier; from 0 with no upper limit; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 22 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Claude Opus 5.5Max1.14#9
2GPT-6 AstraMax1.13#10
3Claude Opus 5Max1.11#26
4Claude Fable 5.1Max1.09#20
5Claude Opus 4.8Max1.01#65
6GPT-5.6 SolMax0.92#37
7GLM-5.3Max0.89#72
8Kimi K3Thinking0.89#61
9GPT-5.5xHigh0.84#54
10Claude Opus 4.6Max0.78#113
11Claude Opus 4.7Max0.75#85
12GPT-5.2High0.7#185
13GPT-5.4xHigh0.6#135
14GPT-5High0.53#178
15O3High0.43#237

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=tasteval-performance-multiplier · How It Works · Data refreshed daily, snapshot 2026-11-21.