Composite-Bench: leaderboard
Composite-Bench evaluates browser-based computer-use agents in deterministic enterprise web environments. The input modality is text and web interaction. Its pinned metric is Macro Success (%), aggregating pass@1 performance on the main slices reported in the source leaderboard. Source: BenchmarkList’s attributed result table.
Metric: Macro Success (%). Source: benchmarklist.com. Saturation forecast: Around September 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (Max) | 74 | #20 (Claude Fable 5) |
| 2 | Claude Opus 4.8 (Max) | 73 | #39 (Claude Opus 4.8) |
| 3 | GLM-5.2 (Max) | 63 | #80 (GLM-5.2) |
| 4 | Kimi K3 | 31 | #35 |
| 5 | Gemini 3.5 Flash (High) | 27 | #66 (Gemini 3.5 Flash) |
| 6 | GPT-5.6 Sol (Max) | 18 | #22 (GPT-5.6 Sol) |
| 7 | Grok 4.5 (High) | 9 | #54 (Grok 4.5) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=composite-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.