Kimi K3 Tech Report - Online Experience: leaderboard
Moonshot AI's in-house Online Experience benchmark mirrors the distribution of real online agent usage, scoring the deliverable file types users request most. Kimi K3 technical report, Table 3; score in percent, higher is better. Kimi K3 ran in Kimi Agent, Claude models and GLM-5.2 in Claude Code, GPT models in Codex. Claude Fable 5's score includes 14 refused tasks.
Metric: Score (%). Source: raw.githubusercontent.com. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Max) | 84 | #22 (GPT-5.6 Sol) |
| 2 | Kimi K3 (Max) | 77.9 | #35 (Kimi K3) |
| 3 | Claude Fable 5 (Max) | 74.2 | #20 (Claude Fable 5) |
| 4 | GPT-5.5 (xHigh) | 73.7 | #33 (GPT-5.5) |
| 5 | Claude Opus 4.8 (Max) | 69.4 | #39 (Claude Opus 4.8) |
| 6 | GLM-5.2 (Max) | 64 | #80 (GLM-5.2) |
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-online-experience · How It Works · Data refreshed daily, snapshot 2026-11-05.