CRAB-Bench: leaderboard
CRAB-Bench (Constraint-based Realistic Agent Benchmark, 2026) gives a tool-using agent 200 trip-booking tasks built from constraint graphs, where flights, a hotel and an attraction must jointly satisfy the user among thousands of distractor options. This board is the pass@1 rate in percent with the RUSE simulator, which plays users with personas and human-grounded behaviors in communication style, information disclosure, clarification and error reaction; a task passes only if both the final databases and the conversation verify. Higher is better.
Metric: Pass rate (%): share of tasks passing both concrete-state (database) and abstract-state (conversation factuality and completion) verification, with the RUSE human-aligned user simulator (personas sampled uniformly, all four behavior dimensions); 200 trip-booking tasks (departure and return flight, hotel, attraction) generated from constraint graphs with node and edge distractors, tau-bench framework, pass@1, default inference and reasoning settings, Claude Sonnet 4.6 simulates the user; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 61 | #406 |
| 2 | GLM-5 | 44 | #260 |
| 3 | Claude Sonnet 4.6 | 32 | #161 |
| 4 | Qwen3 Coder Next | 20 | #714 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=crab-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.