70.2%
#1 of 7
τ-bench pass@1 of Claude Opus 4.5, the highest of the top seven; GPT-5.2 scored 69.9%
What the chart shows
τ-bench places an agent in domains such as retail and airline customer service, with databases, policy rules and a multiturn conversation with a user, and checks whether the end result, often the database state, is correct. Claude Opus 4.5 led at 70.2%, ahead of GPT-5.2 (69.9%) and Qwen3.5-397B-A17B (68.4%).
The top seven span 7.3 percentage points, down to Claude Sonnet 4.5 at 62.9%. Pass@1 counts success on a single attempt, so even the leader fails about three in ten tasks.
Key numbers
| Measure | Value |
|---|---|
| Claude Opus 4.5 | 70.2% |
| GPT-5.2 | 69.9% |
| Qwen3.5-397B-A17B | 68.4% |
| Gemini 3 Flash | 67.8% |
| Claude Sonnet 4.5, 7th | 62.9% |
| Gap, 1st to 7th | 7.3 percentage points |
How to read it
- Scores are from the τ-bench text leaderboard, as cited by the AI Index, which shows the top seven models.
- Success is judged on the final outcome, which is often verifiable from the database state.
Source: τ-bench Leaderboard, 2026, via the AI Index Report 2026 (Stanford HAI), figure 2.6.6: τ-bench: pass@1. Chart and text by AI Adoption.
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Numbers checked against the published data on 20 October 2026.
Download the data (CSV) · Cite this chart
Discussion