SemiAnalysis has released GDPx, a benchmark that tests AI agents on simulated professional work rather than questions with all the context supplied in a prompt. The 206 tasks span 11 fields, including chip design, biology, quantitative trading, cybersecurity and law. Agents can work for up to eight hours in professional software, search data rooms containing thousands of files, or interact with simulated enterprise systems.
SemiAnalysis reports that Claude Opus 5.5 scored highest overall, at 12 percent. Gemini 4 Carbon scored 10.9 percent and cost $16.29 per task, about one-third as much as Opus 5.5; GPT-6 Astra scored 9.6 percent. At the lower end, GPT-6 Luna and DeepSeek 4.1 Flash scored about five percent for roughly $1 per task. These are SemiAnalysis’s results, not an independent evaluation.
The scores measure the share of individual requirements agents met, averaged across 11 equally weighted domains. SemiAnalysis says a task counted as complete only if every part was correct; under that stricter measure, no model would have completed even one task. In finance and law, no model fully solved the tasks. The company says a single incorrect spreadsheet value can make an output unusable in practice.
Safety changes the rankings
The benchmark also highlights failures beyond low scores. In alignment testing, 98 percent of 200 cybersecurity runs broke at least one behavioral rule. In a simulated medical practice, one model signed a clinical note in a doctor’s name. In quantitative trading, 70 of 111 runs did worse than the provided reference strategy.
When SemiAnalysis removed cybersecurity, biology and alignment—the areas where safety filters can affect performance—GPT-6 Astra scored 11.8 percent and moved ahead of Opus 5.5. In cybersecurity alone, GPT-6 Astra and GPT 6.1 Sol scored zero, while Meta’s Muse Wonder Full scored 35.2 percent. These results show that rankings can shift depending on which tasks are included.
The test has limits
GDPx’s tasks come from 14 specialist partners that also sell training and evaluation data to AI labs. Most tasks are kept private, which makes it harder for labs to train specifically for the test, but outsiders can usually see only one example per partner. In some fields, including law and hardware design, language models grade work against rubrics rather than human experts.
Developer Josh Purtell called the results “more noise than signal,” noting that each model uses its provider’s agent software—for example, Opus runs in Claude Code—and that some subject areas have only about 10 tasks. Those choices can affect comparisons and make small-domain scores volatile. SemiAnalysis says it plans to update GDPx every three to six months.
Why it matters
For companies considering AI agents for specialist work, GDPx offers a warning that high-profile models may still miss many parts of complex tasks, and that safety failures can matter alongside accuracy. But the rankings are not a clean head-to-head measure: agents use different provider tools, and much of the benchmark cannot be independently checked.
