AssayBench - Adjusted nDCG@100: leaderboard
Rank the genes a phenotypic screen will call hits, on Genentech's temporal (by-year) split: train on past screens, rank the later ones. Adjusted nDCG at 100, higher is better; only the leaderboard's LLM category is published here. Source: genentech.github.io/AssayBench.
Metric: Adjusted nDCG@100 (%). Source: genentech.github.io. Saturation forecast: Around 2030. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Pro | 15.7 | #82 |
| 2 | Gemini 3.1 Pro (Preview) | 14.72 | #64 |
| 3 | GPT-5.4 | 14.7 | #83 |
| 4 | Gemini 3 Flash | 14.46 | #103 |
| 5 | GPT-5 Mini | 13.62 | #193 |
| 6 | GPT-5.2 | 13.32 | #115 |
| 7 | Claude Opus 4.5 | 12.69 | #87 |
| 8 | GLM-5 | 12.2 | #149 |
| 9 | Claude Sonnet 4.5 | 12.19 | #148 |
| 10 | Kimi K2.5 | 12.08 | #152 |
| 11 | Qwen 3.5 397B A17B | 11.1 | #154 |
| 12 | DeepSeek V3.2 | 10.76 | #210 |
| 13 | Qwen 3 235B A22B 2507 | 10.45 | #341 |
| 14 | GPT-OSS-20B | 10.31 | #523 |
| 15 | Qwen 3.5 122B A10B | 10.12 | #184 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=assaybench-adjusted-ndcg-100 · How It Works · Data refreshed daily, snapshot 2026-11-05.