GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

GDPx is a benchmark covering the most valuable domains for frontier intelligence and aims to be the world's best benchmark for measuring general agentic capabilities. It spans chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering, while intentionally forgoing government, real estate, and insurance. According to GDPx, even the best model scores just 12%.

Results and model strengths

GDPx reports that Opus 5.5 and GPT-6 Astra have sandbagged overall scores due to safety classifiers. OpenAI and Anthropic dominate the top, while SpaceXAI, Meta, and open-weight models lag. Gemini 4 Carbon is the only model near Opus and Astra. Muse Wonder Full leads in cyber. Spikiness is meaningful: Opus 5.5 wins quant trading, hardware design, and chip design; Gemini 4 Carbon wins bio and software engineering; GPT-6 Astra wins RSI and materials science. Finance and legal tasks are so hard that basically every model scores exactly 0. Cost versus performance puts Luna and DeepSeek at the low end, Opus at the high end, and OpenAI in the middle.

How GDPx differs from GDPval

GDPx argues GDPval, released September 25, 2025, is inadequate for today's frontier models because its tasks are non-interactive and spoon-feed context. A finance task asks for up to a 15 page PDF from a roughly 500 word description, lacking a realistic environment. GDPx also cites noisy rubric graders, broken rubrics, incorrect golden solutions, and stale tasks. In contrast, a Living Context finance task is seeded by a real transaction: Bannock Capital buys about 62% of Haverly Snack Company for $880M plus an earnout. The environment has about 4000 anonymized files, and the AI must navigate a deal room of about 2000 files. Cyber, materials, and hardware tasks also use realistic environments.

Methodology and partners

GDPx focuses on domains expected to be most accelerated by AI in the next one to two years, arguing that a benchmark based on share of GDP or the SOC/O*NET occupational taxonomy is a mistake. It partnered with 14 specialized data partners, including Extrapolate for RSI and Beautiful Machines for hardware design. Partners faced rigorous QA and had to pass automated QA on the first try. The default time limit is eight hours. Models ran in native harnesses when available, otherwise OpenCode. Tasks use binary or partial grading, and the headline score is pass@1 across three attempts, with all domains weighted equally.

Deep dives: RSI and hardware design

RSI tasks, provided exclusively by Extrapolate, define RSI as an agent's ability to autonomously build the next version of itself. One task recreates debugging a large-scale RL run where a harness bug injected a message at 32K tokens; agents receive about 1,000 rollout traces comprising 21.2 million tokens. Across 20 rollouts from Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, and GLM 5.3, all four models struggled. Hardware design tasks, provided exclusively by Beautiful Machines, use specs and artifacts from projects such as a 7,000-component robotic patient-transfer system. Rollouts run in sandboxed environments with SOLIDWORKS, Siemens Designcenter NX, Autodesk Fusion 360, KiCad, AMD Vivado, and the Cadence suite, with grading combining deterministic verifiers and LLM-as-a-judge.

Sources: Semianalysis