WebVoyager leaderboard
WebVoyager
18 models tested · Updated 2026-03-26 · Verified sources only
Alumnium MCP + Claude Code leads at 98.5%
1
Alumnium MCP + Claude Code
Alumnium · Blog/Alumnium · 2026-03-26
New WebVoyager SOTA. MCP server + Claude Code general-purpose agent. ~$5 total cost for 610 tasks. Reproducible code available.
98.5%
2
Surfer 2
H Company · Blog/H Company · 2025-10-17
Cross-platform computer-use agent. SOTA on WebVoyager at time of release. Separates planning from execution with orchestrator + sub-agents.
97.1%
3
Gemini 3.1 Pro
Google DeepMind · arxiv/2604.08516 · 2026-04-09
Google computer-use-preview mode on WebVoyager. Tested in MolmoWeb paper.
88.6%
4
Gemini computer-use-preview
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
88.6%
5
GLM-5V-Turbo
Zhipu AI · X/@VaibhavSisinty + Z.ai launch · 2026-04-01
Vision-language model with native multimodal input. Marginal lead over Claude Opus 4.6 (88.0) on browser navigation tasks.
88.5%
6
GUI-Owl-1.5-32B-Thinking
Alibaba · arxiv/2602.16855 · 2026-02-15
Strong web navigation. 32B thinking model for multi-platform GUI tasks.
82.1%
7
SoM Agent + o3
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
79.3%
8
MolmoWeb 8B
Allen AI · Blog/Allen AI · 2026-03-24
Open-weight 8B browser agent. With test-time scaling (pass@4): 94.7%. Outperforms Fara-7B and GPT-4o-based agents.
78.2%
9
MolmoWeb-8B
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
78.2%
10
GUI-Owl-1.5-8B-Thinking
Alibaba · arxiv/2602.16855 · 2026-02-15
Web navigation benchmark. Strong for an 8B model.
78.1%
11
MolmoWeb 4B
AI2 · arxiv/2604.08516 · 2026-04-09
4B open-weight visual web agent. Outperforms all other open-weight models on WebVoyager including 7B models (Fara, Holo1, UI-Tars).
75.2%
12
MolmoWeb-4B
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
75.2%
13
Fara-7B
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
73.5%
14
GPT-5.4
OpenAI · arxiv/2604.08516 · 2026-04-09
OpenAI computer-use-preview mode on WebVoyager. Below MolmoWeb-4B.
70.9%
15
OpenAI computer-use-preview
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
70.9%
16
GLM-4.1V-9B-Thinking
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
66.8%
17
UI-TARS-1.5-7B
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
66.4%
18
SoM Agent + GPT-4o
arxiv · arxiv/2604.08516 · 2026-04-09
Fully open web agents (4B/8B) that outperform similar-scale open models and even SoM agents built on GPT-4o. MolmoWeb-8B achieves 78.2% on WebVoyager and 35.3% on Online-Mind2Web pass@1.
65.1%
benchmark.space canonical public ledger of AI capability