Making AI Headlines · Page 8
AI Agents Enter a Work Benchmark and Leave With Less Than 12 Percent
· New York time
No model scored above 12 percent on SemiAnalysis' new GDPx benchmark, which tests AI agents on simulated knowledge work ranging from chip design to quantitative trading. Claude Opus 5.5 led narrowly over Gemini 4 Carbon. The result offers a contrast to sweeping claims about agents taking over complex office work. GDPx asks models to perform tasks drawn from demanding fields rather than simply answer a general question. On this test, the leading score remained below 12 percent, with little separating the top two models. The benchmark also has limits. Critics point to a small number of tasks in each domain and uneven testing conditions, factors that can affect how much a score says about broader performance. A result on this particular set of simulated tasks is not a complete measure of every workplace use. Even with those qualifications, the figures make a memorable office report. Vendors can describe an agent as tireless, versatile and ready for responsibility, but a difficult assignment still has a way of asking for the work itself. The leaderboard provides a winner and an unusually modest margin for celebration. In the simulated workplace, the machines have discovered that being first in the class is not the same thing as passing the test.
Read this story in the full newspaper →Satire. Not factual reporting. Read the original reporting from the linked publishers.