SemiAnalysis released GDPx on November 2, 2026, a new benchmark for measuring how far Super Intelligence agents can go on real expert work, not just answering questions. Its authors are Max Kan, Andrew Megalaa, and Dylan Patel. Version one holds 206 tasks across 11 fields, built with 14 specialist data partners.
The first results became a talking point immediately, because every score is low. Anthropic's Claude Opus 5.5 leads at 12.0%. Google's Gemini 4 Carbon is second at 10.9%, and OpenAI's GPT-6 Astra third at 9.6%. Meta's Muse Wonder Full is fourth at 9.0%, while Z.ai's GLM 5.3 sits last of the 13 models tested, at 2.7%.
Cost is printed on the leaderboard too, and the comparison is worth a look. One GDPx task costs an average of $51.41 for Opus 5.5 and $53.99 for GPT-6 Astra, but only $16.29 for Gemini 4 Carbon. Gemini trails the leader by just 1.1 points, at roughly a third of the cost. For a team running models every day, that cost figure matters as much as the score.
The tasks are not about fast answers
GDPx is deliberately hard. Each task carries an eight-hour limit and a realistic work setting, not one tidy prompt. Models run in their native harness where available: Claude Code for Opus, Codex for OpenAI's models, Antigravity for Gemini, and OpenCode as the fallback. The headline score is pass@1 across three attempts, every field weighted equally, and grading follows each task's intended method, pass or fail, or partial credit.
The eleven fields are chip design, hardware design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, materials science, alignment, and software engineering. SemiAnalysis says it picked fields where frontier models are expected to earn the most economic value, not by their share of GDP. Government, real estate, and insurance are deliberately left out.
Where each model wins
No single model dominates everything. Opus 5.5 wins quant trading, hardware design, and chip design. Gemini 4 Carbon leads biology and edges software engineering. GPT-6 Astra leads recursive self-improvement and materials science. SemiAnalysis also calls Muse Wonder Full the best available cyber model, by a clear margin.
Two fields are the floor for every model. Under pass-or-fail grading, finance and legal come back at practically zero across the board, so SemiAnalysis reports partial credit for finance. That is not a small technical detail. It means work that demands full accuracy and accountability, like marking up a legal document until it is usable, is still far beyond today's agents.
Two caveats that change how to read the board
The first caveat is about safety filters. SemiAnalysis itself says the Opus 5.5 and GPT-6 Astra scores are held down by their own safety classifiers in cybersecurity, biology, and alignment, because the models sometimes refuse tasks. Strip those three fields out and the board changes: GPT-6 Astra edges Opus 5.5 by 0.1 points, and Gemini 4 Carbon slips to fifth. So the top order depends on which fields count.
The second caveat is about who did the measuring. These are SemiAnalysis's own scores, not independently re-run. GDPx is also a paid product, and its authors have an interest in it becoming the benchmark of record. That does not make the numbers wrong, but an honest reading has to name the source: one credible research firm's claim, not a third-party test result.
GDPx is positioned as a successor to OpenAI's GDPval, released in September 2025 and still the general benchmark that shows up most often on model release cards. The difference is that GDPval hands a model a prompt and a document and asks for a deliverable, while GDPx gives an agent a work environment and hours to work in it. SemiAnalysis plans to update the benchmark every three to six months, so this 12% board is a starting point, not a final verdict.
For context, this lands on the same day as Meituan's LongCat-2.5 release, a model built for coding work that runs long, as in our LongCat article. GDPx tests claims like that in a harder way: not whether a model can answer, but whether it can finish expert work end to end. And if you want the agent that works without being called, compare this with DeepSeek's Super-agent prototype, which is still research code.
Sources: SemiAnalysis announcement (Nov 2, 2026) via Unite.AI (Nov 2, 2026), CryptoBriefing (Nov 2, 2026), SiliconNoon, ModelDex, and OfficeChai (Nov 2, 2026). Figures are snapshots at writing time, evening November 2, 2026 WIB, and can change.
Writing about crypto and Super Intelligence: market news, new projects, Super Intelligence models, and early opportunities like airdrops and testnets. Every piece starts from the official announcement, with the costs worked out up front.