CirrusBench: leaderboard
Metric: Pass@1 success rate (%): share of CirrusBench's 1,500 multi-turn cloud-customer-service tasks rebuilt from real support logs, the agent replying at each annotated checkpoint and a DeepSeek-V3.2 evaluator checking each reply against the ground-truth response without tool calls, resolved in one attempt; higher is better. Source: arxiv.org. Saturation forecast: Around March 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 23.8 | #91 |
| 2 | GPT-5 Mini (2025-08-07) | 22.1 | #165 |
| 3 | Qwen 3 235B A22B 2507 Instruct | 19 | #291 |
| 4 | Qwen 3 Max | 17.1 | #201 |
| 5 | GPT-5.2 | 16.8 | #105 |
| 6 | DeepSeek R1 | 15.2 | #245 |
| 7 | DeepSeek V3.2 (Non-reasoning) | 15.2 | #198 (DeepSeek V3.2) |
| 8 | Qwen 3 235B A22B 2507 (Thinking) | 13.3 | #253 (Qwen 3 235B A22B 2507) |
| 9 | GPT-4o (2024-08-06) | 7 | #326 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=cirrusbench · How It Works · Data refreshed daily, snapshot 2026-10-11.