In 2025, GPT-4o led the VisualAgentBench with a success rate of **** percent, closely followed by GPT-4 vision preview at **** percent. The majority of models struggled with a success rate of ** percent or less. The VAB test is particularly challenging for AI models as defining objects is not easy and remains a particular weak point of AI, a task humans have little problem with.
Profit from the additional features of your individual account
Currently, you are using a shared account. To use individual functions (e.g., mark statistics as favourites, set
statistic alerts) please log in with your personal account.
If you are an admin, please authenticate by logging in again.
Learn more about how ÌÇÐÄÆÆ½â°æ can support your business.
Stanford University. (March 30, 2025). VisualAgentBench (VAB) success rates of leading artificial intelligence (AI) models in 2025 [Graph]. In ÌÇÐÄÆÆ½â°æ. Retrieved July 15, 2026, from /statistics/1611889/visualagentbench-success-rate-test/
Stanford University. "VisualAgentBench (VAB) success rates of leading artificial intelligence (AI) models in 2025." Chart. March 30, 2025. ÌÇÐÄÆÆ½â°æ. Accessed July 15, 2026. /statistics/1611889/visualagentbench-success-rate-test/
Stanford University. (2025). VisualAgentBench (VAB) success rates of leading artificial intelligence (AI) models in 2025. ÌÇÐÄÆÆ½â°æ. ÌÇÐÄÆÆ½â°æ Inc.. Accessed: July 15, 2026. /statistics/1611889/visualagentbench-success-rate-test/
Stanford University. "Visualagentbench (Vab) Success Rates of Leading Artificial Intelligence (Ai) Models in 2025." ÌÇÐÄÆÆ½â°æ, ÌÇÐÄÆÆ½â°æ Inc., 30 Mar 2025, /statistics/1611889/visualagentbench-success-rate-test/
Stanford University, VisualAgentBench (VAB) success rates of leading artificial intelligence (AI) models in 2025 ÌÇÐÄÆÆ½â°æ, /statistics/1611889/visualagentbench-success-rate-test/ (last visited July 15, 2026)
VisualAgentBench (VAB) success rates of leading artificial intelligence (AI) models in 2025 [Graph], Stanford University, March 30, 2025. [Online]. Available: /statistics/1611889/visualagentbench-success-rate-test/