OmniaBench - DAG: leaderboard
Multi-turn tasks in which an agent works through a sampled chain of dependent tool calls in a stateful simulated environment while a simulated user reveals requirements; from OmniaBench (2026). Pass@1 in percent, a task counts only if every rubric item is met; higher is better.
Metric: Pass@1 (%; 354 multi-turn stateful tasks built from sampled tool-dependency graphs, one per level-2 domain, with a simulated user; a task counts as solved only when every rubric item is satisfied (Program tasks by a VerifyCode check); up to 200 tool-call steps; challenging set of 644 tasks evaluated June 29 to July 15, 2026). Source: arxiv.org. Saturation forecast: Around September 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 (Max) | 57.34 | #111 (Claude Sonnet 5) |
| 2 | GPT-5.6 Sol (High) | 55.37 | #30 (GPT-5.6 Sol) |
| 3 | GLM-5.2 (Max) | 54.8 | #138 (GLM-5.2) |
| 4 | GPT-5.5 (High) | 54.8 | #49 (GPT-5.5) |
| 5 | Gemini 3.5 Flash (High) | 54.24 | #101 (Gemini 3.5 Flash) |
| 6 | Claude Opus 4.7 (Max) | 53.39 | #79 (Claude Opus 4.7) |
| 7 | DeepSeek V4 Pro (Max) | 52.54 | #188 (DeepSeek V4 Pro) |
| 8 | DeepSeek V4 Flash (Max) | 51.13 | #258 (DeepSeek V4 Flash) |
| 9 | GLM-5 (Thinking) | 50.85 | #260 (GLM-5) |
| 10 | Qwen 3.5 397B A17B | 50.56 | #246 |
| 11 | Gemini 3.1 Pro (Preview) | 50.28 | #94 |
| 12 | Kimi K2.6 (Thinking) | 49.72 | #178 (Kimi K2.6) |
| 13 | Qwen 3.7 Max (Thinking) | 48.59 | #145 (Qwen 3.7 Max) |
| 14 | Qwen 3.6 27B | 48.02 | #341 |
| 15 | Qwen 3.6 35B A3B | 43.22 | #462 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=omniabench-dag · How It Works · Data refreshed daily, snapshot 2026-11-13.