SemiAnalysis launches GDPx, an agent benchmark where the best model scores 12%

SemiAnalysis released GDPx, a new benchmark that measures how well frontier AI models handle agentic work across 11 economically valuable domains. The firm says it chose fields where it expects frontier AI to pay off most, such as chip design, cybersecurity and biology, and deliberately left out areas like government, real estate and insurance. Its argument is that the first trillion dollars of AI revenue will concentrate in high-return sectors rather than spread evenly across the economy.

According to SemiAnalysis, GDPx is hard: even the best model scores just 12%. The group also says some models, including Opus 5.5 and GPT-6 Astra, get lower overall scores because their safety classifiers sometimes kick in on the cyber, biology and alignment tasks. Leaving those three domains out changes the ranking, and SemiAnalysis says that makes safety classifiers an underrated factor when picking a model. By its testing, Muse Wonder Full is currently the best cyber model by a clear margin.

Each domain was built with at least one specialized data partner, with SemiAnalysis handling final quality checks. Fireworks AI ran inference for open-source models and Modal supplied sandbox infrastructure. A dedicated site shows a live leaderboard, per-domain results and cost per task.

Why it matters

It gives developers and buyers a harder, domain-grounded way to judge which frontier model is actually best for a given high-stakes use case, rather than relying on generic benchmark scores.

Source: SemiAnalysis