New benchmark of AI agents on high-value work finds the best model scores about 12%
This is one firm's first release, and it has not been independently reproduced. Most of the write-up is behind a paywall, so this edition read the open sections and the public leaderboard; the exact score of each model other than the leader was not confirmed. SemiAnalysis says safety filters lower the scores of Opus 5.5 and GPT-6 Astra in cybersecurity, biology and alignment, so the overall ranking partly measures refusals rather than ability. Most task partners are startups that build training environments for AI models, and the sections this edition read do not say whether any tested model was trained on similar tasks. Scores are averaged equally across the 11 fields, which the firm chose; they are not a measure of economic value.
What was claimed?
On November 2, 2026, the research firm SemiAnalysis released GDPx, a benchmark that tests AI agents on 206 tasks across 11 fields: chip design, cybersecurity, biology, finance, law, quantitative trading, AI research, hardware design, materials science, alignment and software engineering. It built the tasks with 14 specialised companies. Agents work inside simulated workplaces for up to eight hours per task, using each developer's own agent software where one exists. SemiAnalysis reports that the best model, Anthropic's Opus 5.5, scores about 12%. It also reports that Opus 5.5 leads in quantitative trading, hardware design and chip design, Google's Gemini 4 Carbon in biology and software engineering, and OpenAI's GPT-6 Astra in AI research and materials science. It says almost every model scores zero on the finance and legal tasks under pass-or-fail grading.
What does the evidence support?
Independent evidence, scope limited
SemiAnalysis argues that the most widely reported benchmark of office work, OpenAI's GDPval from September 2025, hands models all the context in a single prompt. GDPx instead asks agents to work through long, realistic tasks in the fields where AI could earn the most revenue. The low top score shows how far today's models are from doing that work on their own, and the results by field show that no single developer leads everywhere.
What are the limits?
This is one firm's first release, and it has not been independently reproduced. Most of the write-up is behind a paywall, so this edition read the open sections and the public leaderboard; the exact score of each model other than the leader was not confirmed. SemiAnalysis says safety filters lower the scores of Opus 5.5 and GPT-6 Astra in cybersecurity, biology and alignment, so the overall ranking partly measures refusals rather than ability. Most task partners are startups that build training environments for AI models, and the sections this edition read do not say whether any tested model was trained on similar tasks. Scores are averaged equally across the 11 fields, which the firm chose; they are not a measure of economic value.
Claim or evidence date: 2026-11-02. Last verified: .
SemiAnalysis · GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains ↗
Colors classify evidence; they do not rank truthfulness. Explore hype versus reality →