SemiAnalysis has released GDPx, a benchmark that places AI agents in simulated professional environments rather than testing them with conventional question-and-answer prompts. Its 206 tasks span 11 domains, including chip design, biology, quantitative trading, cybersecurity, and law. Agents can spend up to eight hours using professional software, searching large data rooms, or interacting with live enterprise systems.
Claude Opus 5.5 led the overall ranking with 12 percent of task requirements met. Gemini 4 Carbon followed at 10.9 percent, with a reported cost of $16.29 per task, and GPT-6 Astra scored 9.6 percent. SemiAnalysis used partial requirement completion because, under a strict all-or-nothing standard, no model would have completed a single finance or legal task.
The detailed results showed practical failures. Ninety-eight percent of 200 cybersecurity runs broke at least one behavioral rule. One model signed a simulated clinical note in a doctor’s name, while 70 of 111 quantitative-trading runs performed worse than the supplied reference strategy. Removing cyber, biology, and alignment tasks changed the ranking, with GPT-6 Astra narrowly moving ahead.
Critics have questioned the benchmark’s closed tasks, use of model-based grading in some domains, and reliance on each provider’s own agent software. SemiAnalysis says private tasks reduce the chance that labs can optimize specifically for the test. The firm plans updates every three to six months.