SemiAnalysis launched GDPx on November 2, a benchmark built to measure frontier AI models on long, agentic work in the domains its analysts expect to generate the most AI revenue. The headline result is blunt: the best model scores 12%. The report is written by Max Kan, Andrew Megalaa and Dylan Patel.
GDPx covers 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis said it picked domains by expected impact over the next one to two years rather than by share of GDP, and intentionally left out government, real estate and insurance.
The tasks come from 14 specialized reinforcement-learning environment partners, including Datacurve for software engineering, Latch Bio for biology, Chakra for legal and Fleet for alignment. Each task runs inside a simulated environment rather than a single prompt. The finance environment, for example, is a deal room of roughly 4,000 files. Models get up to 8 hours by default.
Key facts
- 01SemiAnalysis published GDPx on November 2, 2026; the best frontier model scores 12% on it.
- 02GDPx spans 11 domains, from chip design and quant trading to legal, finance and alignment.
- 03Tasks run in simulated environments for up to 8 hours, scored pass@1 across three attempts.
- 04Claude Opus 5.5 leads overall; GPT-6 Astra wins recursive self-improvement; Gemini 4 Carbon leads biology.
- 05On one legal redline task, Opus 5.5 averaged 44% partial credit versus 25% for GPT-6 Astra.
Each lab's model ran in its native harness where one exists: Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon. Models without a harness ran in OpenCode. The headline metric is pass@1 across three attempts, with every domain weighted equally.
Claude Opus 5.5 comes out as the best all-around model, winning quant trading, hardware design and chip design. GPT-6 Astra wins recursive self-improvement and materials science. Gemini 4 Carbon is the top biology model and edges ahead in software engineering, and SemiAnalysis called it the only model in the same ballpark as Opus and Astra. The authors named Muse Wonder Full the best available cyber model by a meaningful margin.
Finance and legal proved the hardest. Under binary grading, nearly every model scored exactly 0, so SemiAnalysis also reports partial credit. On a Series B redline task, Opus 5.5 averaged 44%, GPT-6 Astra 25% and Gemini 4 Argon 2%. On a 25-attempt quality-of-earnings task that no model solved, GPT 6.1 Sol averaged 5.6 of 11 checks and Opus 5.5 and GPT-6 Astra averaged 5.4.
The report adds a caveat on safety systems. SemiAnalysis said Opus 5.5 and GPT-6 Astra have sandbagged overall scores due to their safety classifiers, and the ranking changes when cyber, biology and alignment are excluded. On cost against performance, the authors said Luna and DeepSeek compete at the low end, Opus wins the high end and OpenAI claims the middle.
SemiAnalysis positioned GDPx as a successor to OpenAI's GDPval, released in September 2025, which it said remains the most-reported general benchmark on model release cards. The authors called GDPval impressive for its time but said its tasks are fully specified in the prompt, its rubric grading is noisy, with 71% human expert agreement on the gold subset, and some of its reference solutions are wrong.
Quality control was a stated design goal. SemiAnalysis said every partner had to pass automated QA on the first try, and that the same QA process flags nearly 50% of Terminal Bench 4.0 tasks as broken.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
