DuetBench–Diagnosis Elo
Agents
Decagon's DuetBench–Diagnosis arena replays real Duet customer-support diagnosis conversations from the first human turn in Decagon's Duet harness and has a blinded Gemini 3.6 Flash judge compare model responses head to head on task outcome, investigation, tool use, and communication. A Bradley–Terry model converts the pairwise outcomes into Elo ratings centered at 1,000 within the evaluated nine-model pool, and the headline averages the four axis ratings. Pool-relative ratings are kept diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 1035 |
| GLM-5.3 | 1032 |
| Sonnet 5 | 1022 |
| Kimi K3 | 1011 |
| GLM-5.2 | 1009 |
| Opus 4.6 | 998 |
| GPT-5.6 Sol | 990 |
| GPT-5.6 Terra | 952 |
| GPT-5.6 Luna | 951 |