SemiAnalysis has launched GDPx, a new benchmark for general agentic AI, and its first leaderboard carries a striking headline: even the best frontier model scores just 12%. Anthropic’s Claude Opus 5.5 leads the ranking, ahead of Google’s Gemini 4 Carbon and OpenAI’s GPT-6 Astra, in a test designed to measure what AI agents can actually do across eleven high-value domains.
What GDPx measures
GDPx covers chip design, cybersecurity, biology, finance, legal work, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. SemiAnalysis chose those domains not because they are the largest share of the economy, but because it believes the first trillion dollars of AI revenue will concentrate where frontier intelligence delivers the highest return on investment.
The benchmark deliberately leaves out government, real estate, and insurance. SemiAnalysis argues that building a general-purpose test from gross domestic product shares or occupational categories would mislead readers about where frontier models matter most. Instead, GDPx asks whether an agent can complete realistic, multi-step work in the fields where buyers are most likely to pay for it.
What the leaderboard says
Claude Opus 5.5 tops the first GDPx leaderboard at 12.0%, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. All three scores sit far below full task completion. Opus 5.5 wins quant trading, hardware design, and chip design, while Carbon leads biology and software engineering, and Astra leads self-improvement and materials science.
Finance and legal work return the starkest result: SemiAnalysis says its tasks there are so challenging that basically every model scores exactly zero. Safety classifiers also held back some scores, notably for Opus 5.5 and GPT-6 Astra in cybersecurity, biology, and alignment. With fewer refusals, Muse Wonder Full is the best available cyber model, SemiAnalysis says.
How GDPx differs from GDPval
GDPx succeeds OpenAI’s GDPval, released September 25, 2025, which SemiAnalysis calls impressive for its time but inadequate for today’s agents. GDPval tasks were precisely specified and non-interactive, with context spoon-fed in the prompt. An agent today must instead search, gather context, use tools, and act inside a working environment.
To build that kind of test, SemiAnalysis partnered with fourteen specialised environment startups and put each partner through quality checks for consistency, completeness, cheating, and unfair verifiers. Tasks run in simulated work environments, such as cybersecurity services, chip-design tools, and materials-science laboratories, with a default eight-hour limit and headline scores reported as pass at one across three attempts.
Our read
The headline number is less a ranking than a reality check. A 12% top score does not mean frontier models are useless; it means GDPx was built to stay hard, requiring expertise, long horizons, and messy environments rather than one polished answer. Buyers who treat a win as proof an agent can replace a team will misread this test.
The spikiness across domains is the more useful signal. No model wins everywhere, and leaders trade places by field, so buyers should choose by task, not brand. Carbon leads biology and software engineering, Astra leads self-improvement and materials science, and Opus 5.5 leads trading and hardware work: different tools for different jobs, not one default.
Why It Matters
Benchmarks shape buying decisions, hiring plans, and investment theses, so a credible new yardstick for agents could shift how companies judge progress. If GDPx gains traction, vendors will face pressure to show performance on long, environment-based tasks rather than short question-and-answer tests. That would reward systems that can plan, recover from errors, and use tools reliably over hours.
The test also raises the bar for benchmark makers themselves. SemiAnalysis says it dropped prominent partners that failed its quality checks first time, and will update GDPx every three to six months. Next, other labs must decide whether to accept its task design, publish on the same terms, and let researchers probe where tasks are fair, flawed, or too hard.
Background
SemiAnalysis is a semiconductor and AI research firm known for detailed analysis of chips, compute, and frontier models. Its new benchmark arrives as agents move from answering prompts to carrying out multi-step work, and as buyers look for evidence beyond marketing claims about what those agents can actually finish.
Source: SemiAnalysis GDPx post and Crypto Briefing report
0 Comments