OneMillion-Bench - Expert Score: leaderboard
Metric: Expert score (%): per task, the weighted share of expert-written rubric points earned (rubric weights from -20 to 10, negative rubrics penalizing violations; each task clipped to 0-1), averaged over tasks on the 200-task Global set (Finance, Law, Healthcare, Industry and Natural Science, 40 tasks each) of OneMillion-Bench's expert-curated professional tasks, vanilla models without tools (deep-research agents search by design and the paper prints the same result in both settings); rubrics graded by an LLM judge; higher is better. Source: arxiv.org. Saturation forecast: Around April 2028. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (High) | 55 | #76 (GPT-5.4) |
| 2 | Claude Opus 4.6 (High) | 55 | #60 (Claude Opus 4.6) |
| 3 | GPT-5.3 Codex (High) | 48.1 | #68 (GPT-5.3 Codex) |
| 4 | GPT-5.2 (High) | 47.6 | #105 (GPT-5.2) |
| 5 | Kimi K2.5 (High) | 45.1 | #139 (Kimi K2.5) |
| 6 | Gemini 3 Pro (Preview) | 43.7 | #64 |
| 7 | DeepSeek V3.2 Speciale | 40.5 | #125 |
| 8 | Doubao-Seed-2-0-Pro-260215 (High) | 39.4 | |
| 9 | Grok 4 | 38.6 | #169 |
| 10 | GLM-5 | 38.5 | #137 |
| 11 | MiniMax-M2.1 | 34.9 | #283 |
| 12 | Step 3.5 Flash | 34.4 | #262 |
| 13 | MiniMax-M2.5 | 29.5 | #295 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=onemillion-bench-expert-score · How It Works · Data refreshed daily, snapshot 2026-10-11.