GDPx Benchmark Debuts: Strongest Model Scores Only 12%, Finance and Legal Tasks All Zero

SemiAnalysis released the GDPx benchmark on November 2, 2026. The strongest model, Claude Opus 5.5, scored only 12.0%, Gemini 4 Carbon scored 10.9%, and GPT-6 Astra scored 9.6%. Under strict binary scoring, all models received 0 on finance and legal tasks.

Factual Reconstruction

According to the GDPx benchmark released by SemiAnalysis, the tests cover 11 domains: chip design, biology, finance, law, quantitative trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. The benchmark was completed through collaboration with 14 specialized RL environment startups, including Aria, Arcverex, Ergodex, Arithmetic, Latch Bio, Living Context, Chakra, Third Moment, Extrapolate, Beautiful Machines, ValenceForge, Emulated, Fleet, and Datacurve. Tasks in each domain underwent rigorous quality control to ensure that scores reflect true capability rather than grader flaws.

The results show that even leading models score only 12% overall. After excluding domains such as cybersecurity, biology, and alignment that may trigger safety classifiers, the rankings adjust somewhat, but finance and legal tasks remain at 0 across the board. Opus 5.5 performs strongly in quantitative trading, hardware design, and chip design; Gemini 4 Carbon leads in biology and software engineering; and GPT-6 Astra wins in recursive self-improvement and materials science.

Mechanism Breakdown

GDPx is designed to focus on domains with high GDP contribution, avoiding sectors such as government, real estate, and insurance, and aims to track the agentic capabilities of frontier models. Tasks require models to make judgments under real-world constraints rather than merely perform arithmetic or pattern matching. The zero scores on finance and legal tasks indicate that although models can handle calculations, they cannot anticipate counterparties' strategies or make complex decisions.

Safety classifiers affect scores for some models; for example, Opus 5.5 and GPT-6 Astra are restricted in certain domains. The benchmark also observes pronounced spikes in model performance, with similar patterns appearing within different model families. The cost-performance Pareto frontier shows Luna and DeepSeek at the low end, Opus at the high end, and OpenAI in the middle; the emergence of Gemini 4 Carbon is worth watching.

Industry Impact

The benchmark directly challenges the narrative that "AI can replace knowledge workers." Models may score relatively high in sandbox environments, but they fail comprehensively when placed in real legal or financial scenarios. This resembles the historical problem of benchmark saturation: early task designs often failed due to grader flaws, while GDPx avoided such problems through collaboration with specialized partners.

For model selection, the importance of safety classifiers is underestimated; for example, Muse Wonder Full performs strongly in cybersecurity. Industry needs to reassess deployment strategies, especially in high-value domains, where relying solely on general benchmarks is no longer sufficient to reflect actual ROI.

Strategic Assessment

[Analysis] The release of GDPx may prompt labs to prioritize optimizing agentic judgment capabilities rather than simply expanding knowledge coverage. OpenAI and Anthropic hold an advantage in the rankings, while SpaceXAI, Meta, and open-source models lag behind, which may influence future investment and collaboration directions. The progress of Gemini 4 Carbon shows Google's potential in specific domains, but its actual performance after official release still needs to be observed overall. Over the long term, such benchmarks can help concentrate resources in high-return domains rather than spreading them across the entire economy.