GPT-5.5 and Claude Fable 5 Lag Behind Humans in ActiveVision Test
AI models fall short of human performance in visual perception tests.
1What is the ActiveVision Test?
ActiveVision is a benchmark designed to measure the visual perception capabilities of AI models. It consists of 17 different tasks and aims to assess repeated visual perception rather than a single static description.
2GPT-5.5's Performance
GPT-5.5 solved only 10.6% of the tasks in the ActiveVision test. It scored zero on 11 out of the 17 tasks, highlighting its limitations in visual perception.
3Claude Fable 5's Results
Claude Fable 5 performed even worse than GPT-5.5, solving only 3.5% of the tasks. Despite leading in reasoning and coding benchmarks, it struggled with visual perception.
4Human Performance
Three human participants successfully completed 96.1% of the tasks in the ActiveVision test. These results underscore the gap between AI models and human visual perception capabilities.
5Model's Ability to Write Its Own Code
The research noted that AI models couldn't patch their shortcomings by writing their own code. This indicates fundamental limitations in their visual perception abilities.
6AI's Development Journey
The ActiveVision test results emphasize that AI models still need significant development in visual perception. More research and innovation are required to reach human-level performance.