SemiAnalysis launches GDPx, where the best model scores 12%
The new agent benchmark covers 206 tasks in 11 domains, including finance and quant trading. Claude Opus 5.5 leads at 12.0%, with Gemini 4 Carbon at 10.9% for about a third of the cost.
Source published · In the Daily Brief of 3 November 2026
Filed under Agents, AI industry
GDPx was built with 14 data partners and runs models in realistic environments with a default limit of eight hours a task. The headline score averages pass@1 over three attempts and weights every domain equally. Claude Opus 5.5 cost $51.41 a task, Gemini 4 Carbon $16.29 and GPT-6 Astra, third at 9.6%, $53.99.
SemiAnalysis presents the benchmark as a successor to OpenAI’s GDPval, whose tasks it calls non-interactive. It said safety classifiers depress some models’ scores. Ethan Mollick of the Wharton School, writing on X, welcomed the benchmark but said it needs human baselines and that its zero scores hide differences in success rates.