GDPx Benchmark: Best AI Model Scores Just 12% on Professional Work
A new benchmark is trying to answer a harder question than most AI tests: can an agent actually do work that the economy pays for? SemiAnalysis launched GDPx on 2 November 2026, scoring AI agents on economically valuable work across 11 domains including chip design, cybersecurity, biology, legal work, quantitative trading, materials science and software engineering, according to SemiAnalysis’s launch post of 2 November 2026.
- SemiAnalysis launched GDPx on 2 Nov 2026 across 11 economically valuable domains
- Anthropic’s Opus 5.5 led overall at just 12% in SemiAnalysis’s own scoring
- The result comes from one firm’s tasks and rubric, with no independent reproduction yet
The headline number is deliberately uncomfortable. By SemiAnalysis’s own scoring, Anthropic’s Opus 5.5 led overall at just 12%, according to the the same SemiAnalysis post. That does not mean everyday automation is failing. It means frontier expert work — the kind that normally requires years of training — is still far beyond what agents can finish unaided.
The design of the test matters. Most benchmarks ask exam-style questions. GDPx instead tries to test jobs, which is why its domains range from hardware to law. The low ceiling is also a reminder of how quickly the wider AI race is moving financially, as seen in our coverage of Anthropic’s reported IPO investor interest, where expectations are running far ahead of proven autonomous performance.
Readers should hold the 12% figure loosely. The source itself cautions that these are one firm’s tasks and one firm’s rubric, with no independent reproduction yet, and that SemiAnalysis sells research to the same industry it is grading. Part of the low scoring, the report notes, may also reflect safety classifiers refusing to engage in some domains — so the benchmark partly measures willingness to attempt a task, not only ability.
Why this matters for buyers
For businesses, the practical lesson is not to abandon agents, but to evaluate them on your own work. A model that struggles with chip design may still draft reports, summarise documents or reconcile simple lists reliably. The gap between benchmark theatre and workplace usefulness is exactly what GDPx exposes. That pressure is visible across the industry, from AI labs to chip suppliers such as those in our report on Samsung’s record profit and sold-out HBM supply, where demand for AI hardware continues to outrun what software can yet deliver on its own.
Public reaction
Reaction in the source coverage was split in a familiar way: investors can read a 12% ceiling as enormous room to grow, while safety advocates can read it as proof that systems are not ready for unsupervised expert work. Both readings use the same number, which is why independent replication will matter more than the launch headline.
What this means
GDPx matters because it reprices the central claim of the agent era. Labs and vendors sell AI agents as ready to take on professional work; SemiAnalysis built a benchmark specifically to test that claim on economically valuable tasks — 206 tasks across 11 domains, from chip design and quant trading to legal work and biology, built with 14 specialist data partners, with task time limits running to eight hours. The result is a ceiling, not a slope: Anthropic’s Opus 5.5, the best overall performer, scored 12%. In finance and legal, essentially every model scored zero. Whatever else changes, procurement conversations in 2027 will happen in the shadow of that number — with the important caveat, noted in our original report, that this is one firm’s own scoring with no independent reproduction yet.
The comparisons inside GDPx are more informative than the headline. SemiAnalysis built the benchmark as an explicit challenger to OpenAI’s GDPval, which it calls inadequate for evaluating today’s agents, so part of what is being contested is who gets to grade the industry. Among models, the ranking is unstable in instructive ways: Gemini 4 Carbon, at 10.9%, is the only model SemiAnalysis places in the same ballpark as Opus 5.5 and GPT-6 Astra, although Carbon is not yet generally available; and the scores of Opus 5.5 and Astra are pulled down by their own safety classifiers in cybersecurity, biology and alignment — excluding those domains reshuffles the order, and SemiAnalysis itself says Meta’s Muse Wonder Full is actually the best available cyber model by a meaningful margin. The cost picture cuts across the quality one: on the launch’s cost-performance frontier, cheaper models fight at the low end while Opus wins the high end.
For US readers — above all the US businesses buying agents now — the disciplined reading is twofold. First, treat 12% as an argument for human-in-the-loop deployment in exactly the high-value domains GDPx tests, not as a verdict on the routine drafting, summarising and reconciling work most companies actually automate, which this benchmark deliberately does not resemble. Second, notice what the benchmark’s existence says about the market: an independent research firm can now build a business grading the graders’ customers. What to watch is the trajectory — future GDPx versions will show whether scores climb at the pace the labs’ roadmaps imply, and whether independent groups reproduce the methodology.
For a US buyer comparing vendors, that instability is itself useful information. A single headline ranking hides three separate questions that GDPx pulls apart: can the model do the work, will its safety classifiers let it attempt the work in domains such as cybersecurity and biology, and what does performance cost on the cost-performance frontier SemiAnalysis published? Those questions will not have the same answer in every department. A legal or finance team reading scores at effectively zero faces a different decision from a software team weighing a cheaper low-end model against Opus at the high end, and both face a different decision again from a security team told that the best available cyber model is Muse Wonder Full. Testing a vendor on a shortlist of your own tasks, with your own reviewers, remains the only way to turn a general benchmark into a purchasing decision.
Read Next
- Anthropic IPO Japanese investors report
- Anthropic Warns AI Could Soon Speed Up AI Development Beyond Evaluation
Sources
