SemiAnalysis has released GDPx, a new benchmark designed to measure what AI agents can actually do at real professional work — and the headline result is a cold shower for the industry. The best model in the world, Anthropic's Claude Opus 5.5, completed just 12.0% of the tasks. Google's Gemini 4 Carbon came second at 10.9%, and OpenAI's GPT-6 Astra third at 9.6%. No current model, the research firm says, comprehensively solves tasks across the domains tested.
Published on November 2 and authored by Max Kan, Andrew Megalaa and Dylan Patel, GDPx is SemiAnalysis's attempt to build the definitive yardstick for general agentic capability: systems that take multi-step actions toward a goal, rather than simply answering questions. Its first leaderboard ranks 13 models, with full results public at gdpx.semianalysis.com.
A workplace test, not a quiz
GDPx version 1.0 contains 206 tasks across 11 domains SemiAnalysis chose for their economic return on investment rather than their share of GDP: chip design, hardware design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, materials science, alignment and software engineering.
Each domain was built with a specialist partner — 14 in total, including Latch Bio for biology, Datacurve for software engineering and Chakra for legal work. Tasks carry an eight-hour limit, and models run in their native harness — Claude Code for Opus, Codex for Astra, Antigravity for Gemini. The headline score averages pass-at-first-attempt results across three attempts, weighting every domain equally.
The tasks are designed to feel like work, not exams. The finance tasks reconstruct a real private-equity deal — a fund buying about 62% of a family snack manufacturer for $880 million plus an earnout — spread across 4,000 anonymised files. The agent must work out which documents matter before it can do anything at all.
Where the models shine — and where they score zero
Underneath the headline, no single model dominates. Opus 5.5 won quant trading, hardware design and chip design outright. Gemini 4 Carbon led in biology and software engineering, while GPT-6 Astra took first place in recursive self-improvement and materials science.
Finance and legal are the floor. Under the benchmark's binary grading — the task is either completed end to end or it is not — every model scored zero in both domains, which is why SemiAnalysis reports partial-credit scores there instead. In legal work, not one redline produced across 20 rollouts and four models was usable.
Elsewhere, 70 of 111 quant-trading runs did worse than the reference strategy, and 196 of 200 alignment cyber runs broke at least one behaviour rule. Cost also varies: Opus 5.5 averaged $51.41 per task, against $16.29 for Gemini 4 Carbon.
A cold shower, not a verdict
SemiAnalysis presents GDPx as an answer to OpenAI's GDPval, published in September 2025, which it argues belongs to the pre-agent era: its tasks hand the model a prompt and a document and ask for a deliverable. GDPx, by contrast, drops the agent inside a workplace and grades whether the job got done.
Two caveats deserve weight. GDPx is a paid SemiAnalysis product whose partner-built tasks have not been independently audited. And a 12% top score reflects tasks chosen to be hard, not proof that models cannot do useful work; the score also measures the harness, not just the model.
Still, GDPx reframes the debate: the question is no longer whether a model can pass a test, but whether it can finish a shift — and on that measure, even the best agent completes one task in eight. SemiAnalysis plans updates every three to six months, so the next reading will show whether that number climbs.
