SemiAnalysis Releases GDPx, an Agentic Benchmark Across High Value Domains Where the Best Model Scores Just 12%
Today we鈥檙e releasing GDPx, a benchmark for measuring agentic capabilities across the most valuable domains. GDPx is a true frontier-level benchmark: even the best model scores just 12% (1/7)馃У
Context
A hard new benchmark built around the domains where AI may earn the most money gives labs and buyers a fresh way to compare frontier agents, with lots of room left to improve.
SemiAnalysis is a research firm best known for its work on chips and AI infrastructure. GDPx is its new benchmark for agentic AI, built with outside data partners across 11 domains. The leaderboard image in the launch post averages each model across all domains, weighted equally.
Deep dive
SemiAnalysis (@SemiAnalysis_) announced GDPx on the evening of Nov 1 ET, describing it as a benchmark for measuring agentic capabilities across the most valuable domains. In its words, it is a true frontier level benchmark, since even the best model scores just 12%.
The leaderboard in the launch post ranks 13 of 20 models. Claude Opus 5.5 from Anthropic leads at 12.0%, followed by Google's Gemini 4 Carbon at 10.9%, OpenAI's GPT 6 Astra at 9.6% and Meta's Muse Wonder Full at 9.0%. Further down are Claude Fable 5.1, GPT 6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon, with DeepSeek 4.1 Flash the best open model at 5.4%. Grok 4.7 scores 3.9% and GLM 5.3 2.7%. The chart also lists cost per task, from $0.92 for GPT 6 Luna to $67.16 for Claude Fable 5.1, with Gemini 4 Carbon at $16.29 against $51.41 for Opus 5.5.
In the thread, SemiAnalysis says each domain was built with specialized data partners, 14 in total, and that it did the final quality checks itself. Fireworks AI provided inference for open models and Modal provided sandboxes. It says GDPx focuses on domains with high return, such as chip design, cyber and bio, rather than weighting by share of GDP.
SemiAnalysis also says the scores of Opus 5.5 and GPT 6 Astra are held down by their safety classifiers, so it also reports results that exclude cyber, biology and alignment. It calls Muse Wonder Full the best available cyber model by a meaningful margin. Google CEO Sundar Pichai quoted the launch, pointing to the cost line for Carbon. These results come from SemiAnalysis and have not been independently reproduced.
What to expect next
Watch whether labs start citing GDPx scores in their own launches, and whether SemiAnalysis adds the other models in its full list. Its point about safety classifiers lowering scores could spark debate over how benchmarks should treat refusals. Expect the cost per task column to get as much attention as the scores, since Google is already pointing to it. Independent reruns of the tasks will show how stable the rankings are.