Benchmarks / Agentic Coding
τ²-Bench
Agentic tool-use benchmark (dual-control customer-service domains). Reported per model by evaluators running the standard τ² harness.
maintainer unknown · 294 observations on record
Versions
1 version
| Version | Tasks |
|---|---|
| τ²-Bench current tau2-bench | — |
Current leaderboard
τ²-Bench · Accuracy · top 25 of 217 systems
| # | System | Accuracy % |
|---|---|---|
| 1 | GLM 5.2max effort Z.ai | 99.1% |
| 2 | GLM 4.7 Flashthinking zai | 98.8% |
| 3 | GLM 5v Turbothinking zai | 98.5% |
| 4 | GLM 5thinking Z.ai | 98.3% |
| 5 | GLM 5.1thinking Z.ai | 97.7% |
| 6 | Grok 4.3high effort xAI | 97.7% |
| 7 | Qwen3 6 Plus Alibaba | 97.7% |
| 8 | GLM 5no reasoning Z.ai | 97.4% |
| 9 | GLM 5.1no reasoning Z.ai | 97.1% |
| 10 | Deepseek V4 Promax effort DeepSeek | 96.2% |
| 11 | Kimi K2 5thinking Moonshot AI | 95.9% |
| 12 | GLM 4.7thinking Z.ai | 95.9% |
| 13 | Kimi K2 6 Moonshot AI | 95.9% |
| 14 | Qwen3 5 397B A17bthinking Alibaba | 95.6% |
| 15 | Gemini 3.5 Flashmedium effort Google DeepMind | 95.6% |
| 16 | Gemini 3.1 Pro Preview Google DeepMind | 95.6% |
| 17 | Deepseek V4 Flashhigh effort DeepSeek | 95.6% |
| 18 | Gemini 3.5 Flashhigh effort Google DeepMind | 95.3% |
| 19 | Minimax M2 5 MiniMax | 95.3% |
| 20 | Mimo V2 Pro xiaomi | 95.0% |
| 21 | Mimo V2 Flashthinking xiaomi | 95.0% |
| 22 | Deepseek V4 Flashmax effort DeepSeek | 95.0% |
| 23 | Qwen3 7 Max Alibaba (Qwen) | 94.7% |
| 24 | Deepseek V4 Flashno reasoning DeepSeek | 94.4% |
| 25 | Claude Opus 4.8max effort Anthropic | 94.4% |
One row per evaluated system — reasoning-effort variants rank separately. Where a system has attr variants for this version (eval splits, style control), only the canonical variant is ranked: split=semi_private, style_control=true, or the unsplit run.