AI Agents Stuck Below 12% on Real-World Task Benchmark

A new benchmark from the analyst firm SemiAnalysis, called GDPx, puts different AI agents into real-world scenarios. Each agent gets tasks to finish within eight hours. The benchmark covers 206 distinct tasks across eleven different industries, including semiconductor design, stock trading, and legal services. No model managed to score above 12% on fully completed tasks. The winner was Anthropic's Claude Opus 5.5, with a score of exactly 12%. Google's Gemini 4 Carbon came in second, at 10.9%. OpenAI's GPT-6 Astra followed, at 9.6%. Finally, the budget Chinese model DeepSeek 4.1 Flash scored around 5%. Not a single model fully completed any task in the finance or legal categories.

This suggests AI agents still cannot reliably handle complex, multi-step work without a human watching over them. One developer criticized the benchmark, saying it is mostly noise and that there are not enough tasks per field to draw real conclusions. Still, the results match a pattern we have already seen: Anthropic leads, while Google and OpenAI trail behind.