MAG (Set-of-Mark): leaderboard
MAG (Multimodal Action and Guide, 2026) is a web-agent benchmark on the six self-hosted WebArena sites: from screenshots, an agent completes a live multi-step task and writes at every step the guide sentence a user would need. This board is the task success rate in percent on 174 held-out test tasks with Set-of-Mark element selection and a 25-step budget; higher is better.
Metric: Task success rate (%; functional WebArena checkers, LLM judge for 29 answer tasks) with Set-of-Mark element selection grounding, 174 held-out test tasks on the six self-hosted WebArena sites; the agent sees screenshots, acts and writes a guide sentence at every step, 25-step budget, one run, the MAG harness; higher is better. Source: arxiv.org. Saturation forecast: Around August 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (Low) | 35.6 | #49 (GPT-5.5) |
| 2 | Gemini 3.5 Flash (Low) | 34.5 | #101 (Gemini 3.5 Flash) |
| 3 | Claude Sonnet 4.6 (Low) | 27.6 | #161 (Claude Sonnet 4.6) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=mag-set-of-mark · How It Works · Data refreshed daily, snapshot 2026-11-13.