WebHarbor - GUI - Agent Judge: leaderboard
WebHarbor Table 8,page 9:306 held-out tasks; ReAct-style GUI vs OpenCode equipped with GUI tools, separate configurations. Verifier and agent-judge pass rates separately reported; their agreement is not capability. General-model effort unspecified. Browser 1280 x 1000,temperature 0.6,top-p 0.95,top-k 20,30-action limit except WebArena 50. Qwen 3-VL-4B-Thinking is the Base row per training setup; trained OpenWebRL and WebHarbor checkpoints remain distinct. Metric: Pass rate (%).
Metric: Pass rate (%). Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra | 95.4 | #6 |
| 2 | Qwen 3.6 35B A3B | 43.4 | #238 |
| 3 | GPT-4.1 | 37.9 | #258 |
| 4 | GPT-5.5 | 27.4 | #33 |
| 5 | Qwen 3 VL 4B (Thinking) | 24.5 | #548 |
No result here: #3 Claude Opus 5.5, #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=webharbor-gui-agent-judge · How It Works · Data refreshed daily, snapshot 2026-11-05.