Kimi K3 Tech Report - Agent Behavior Bench: leaderboard
Moonshot AI's in-house Agent Behavior Bench scores process quality alongside task completion, including tool-use behavior, efficiency and discipline; all models run in the Kimi Work harness. Kimi K3 technical report, Table 3; score in percent, higher is better. Claude Fable 5 had 6 refusals out of 95 tasks.
Metric: Score (%). Source: raw.githubusercontent.com. Saturation forecast: Around February 2027. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 76.4 | #22 (GPT-5.6 Sol) |
| 2 | Claude Fable 5 (Max) | 75.5 | #20 (Claude Fable 5) |
| 3 | GPT-5.5 (xHigh) | 70.1 | #33 (GPT-5.5) |
| 4 | Claude Opus 4.8 (Max) | 65.7 | #39 (Claude Opus 4.8) |
| 5 | Kimi K3 (Max) | 65 | #35 (Kimi K3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kimi-k3-tech-report-agent-behavior-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.