SemiAnalysis launches GDPx: the best AI agent completes just 12% of its high-value tasks, and finance and legal score zero
What's the deal? SemiAnalysisDealroom has a profile for this one. Try Dealroom →, the chip and AI research firm led by Dylan Patel, published GDPx on 2 November, a benchmark for AI agents made of 206 long tasks across 11 domains: chip design, hardware design, materials science, cybersecurity, biology, alignment, finance, legal, quant trading, recursive self-improvement and software engineering. The authors argue the first trillion dollars of AI revenue will concentrate in sectors that can deploy frontier models at high return, so they deliberately left out industries such as government, real estate and insurance. Tasks were built with 14 specialised data partners, among them Arcverex, Aria, Arithmetic, Beautiful Machines, Chakra and Datacurve, with SemiAnalysis doing final quality checks.
The leaderboard: Anthropic's Claude Opus 5.5 leads at 12.0%, ahead of Google's Gemini 4 Carbon (10.9%) and OpenAI's GPT-6 Astra (9.6%). Meta's Muse Wonder Full scores 9.0%, Grok 4.7 3.9% and the open-weight GLM 5.3 2.7%. Gemini 4 Carbon, not yet generally available, gets there at $16.29 per task against $51.41 for Opus. Results are spiky: Opus wins quant trading, hardware and chip design; Carbon leads biology and narrowly software engineering; Astra wins recursive self-improvement and materials science. On finance and legal, "basically every model scores exactly 0".
The caveats: Tasks run for up to eight hours in each model's own harness (Claude Code, Codex, Antigravity, or OpenCode where none exists), scored as pass@1 over three attempts. SemiAnalysis says safety classifiers sandbag Opus 5.5 and GPT-6 Astra in cyber, biology and alignment, and calls Muse Wonder Full "the best available cyber model today by a meaningful margin". These are SemiAnalysis's own runs and choice of domains, not an independent test. Patel's verdict on X: "Official apology to Google from me, they cooked with Carbon."
The signal: Even the best agents fail most long, expert-level tasks, and who leads depends on which domains you count. For buyers, domain scores and cost per task say more than one headline number.
Read more: SemiAnalysis, GDPx leaderboard, SemiAnalysis (X), Dylan Patel (X)
Image credit: SemiAnalysis / GDPx
