ClawBench benchmark: Claude, GPT and Gemini agents fail 33% of tasks on real websites

A new benchmark called ClawBench is drawing attention after tests showed leading AI agents collectively failing on real websites, with an overall failure rate of 33%.

According to a report from Sina, ClawBench was designed to measure how AI agents perform in genuine web environments rather than on optimized, curated test sets. Agents based on Claude, GPT and Gemini — the flagship models from Anthropic, OpenAI and Google — were among those evaluated, and their real-world results fell well short of their scores on traditional benchmarks.

A failure rate of roughly one in three means that even with the most advanced models wired into a browser or workflow, a significant share of everyday operations still cannot be automated reliably. The results reinforce a familiar pattern: agents shine on carefully constructed evaluation sets, but their capabilities shrink markedly when faced with the messy, unpredictable structure of real websites.

For enterprises and developers, the findings are a reminder to validate agent performance in real production scenarios rather than relying on benchmark scores alone before deployment.

What to watch next is whether ClawBench's task design becomes a reference for the industry, and whether model makers respond with optimizations aimed at real-world web use.