T1-Bench: leaderboard

Metric: Pass@3 (%): share of conversations solved in at least one of three trials, where a conversation passes when every ground-truth tool call and parameter is reproduced and every ground-truth tool output is matched; T1-Bench: 525 multi-turn customer-service conversations over 25 domain settings (11 single domains and 14 combinations of 2 to 11 domains, 76 tools), a Gemma4-26B-A4B-it user simulator, three trials per conversation at temperature 1.0; averages weighted by conversations per setting; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 12 models tracked.

Top models

#ModelScore
1Gemma 4 31B (IT)61.33
2Gemma 4 26B A4B (IT)58.86
3Gemma 4 E2B (IT)56
4GPT-OSS-120B (High)55.05
5gemma-4-E4B-it54.67
6Claude Sonnet 4.552.57
7Claude Opus 4.650.67
8Claude Haiku 4.550.67
9GPT-5.4 (Low)36.95
10GPT-OSS-20B (High)36.57

Interactive version: theaggregate.ai/benchmark?slug=t1-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.