SemiAnalysis, the semiconductor and AI research firm, released a new benchmark called GDPx on November 2, 2026. Its first version contains 206 agentic tasks across 11 professional fields and ranks 13 AI models. Anthropic's Claude Opus 5.5 placed first with a score of 12.0%. (SemiAnalysis (opens in a new tab); Unite.AI (opens in a new tab))
Google's Gemini 4 Carbon finished second at 10.9% and OpenAI's GPT-6 Astra was third at 9.6%, according to Unite.AI and OfficeChai. Meta's Muse Wonder Full placed fourth at 9.0%, and Z.ai's GLM 5.3 was last of the 13 models at 2.7%. (Unite.AI (opens in a new tab); OfficeChai (opens in a new tab))
What the benchmark measures
The 11 fields are chip design, cybersecurity, biology, finance, legal work, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. The authors, Max Kan, Andrew Megalaa and Dylan Patel, wrote that they picked fields by how much value they expect AI to create there, not by each field's share of economic output. Government, real estate and insurance were left out on purpose. (SemiAnalysis (opens in a new tab); Unite.AI (opens in a new tab))
Tasks were built with at least one specialized data partner per field, 14 partners in total, and each task ships with a realistic environment instead of a short prompt. The 24 finance tasks, for example, recreate a private equity deal with roughly 4,000 anonymized files. Tasks have a default time limit of eight hours, and models run in their own tools where available: Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon. (SemiAnalysis (opens in a new tab); Unite.AI (opens in a new tab))
The headline score averages each model's pass@1 result over three attempts, with every field weighted equally. SemiAnalysis called GDPx a true frontier benchmark because even the best model reaches only 12%. The firm plans new versions every three to six months. (SemiAnalysis (opens in a new tab); Unite.AI (opens in a new tab))
Where the models struggled
Finance and legal tasks proved so hard under binary grading that every model scored zero, so SemiAnalysis also reported partial scores. On legal partial scores, Opus 5.5 averaged 44% and GPT-6 Astra 25%. In quant trading, 70 of 111 graded runs did worse than the reference strategy, and none of the 12 attempts at one hardware design task produced a verified full pass. (SemiAnalysis (opens in a new tab); Unite.AI (opens in a new tab))
Results varied widely by field. SemiAnalysis named Opus 5.5 the best all-around model, with wins in quant trading, hardware design and chip design. Gemini 4 Carbon was the strongest biology model, GPT-6 Astra led recursive self-improvement and materials science, and Meta's Muse Wonder Full was described as the best cyber model by a meaningful margin. (SemiAnalysis (opens in a new tab); Unite.AI (opens in a new tab))
Cost also differed sharply. Opus 5.5 cost $51.41 per task and GPT-6 Astra $53.99, while Gemini 4 Carbon cost $16.29, according to Unite.AI. (Unite.AI (opens in a new tab); OfficeChai (opens in a new tab))
The safety filter effect
SemiAnalysis said the overall scores of Opus 5.5 and GPT-6 Astra are held down by their safety classifiers in cybersecurity, biology and alignment tasks. With those three fields removed, GPT-6 Astra edges Opus 5.5 by 0.1 percentage points and Gemini 4 Carbon drops from second to fifth. The firm said this adjusted ranking better matches its own assessment of the models. (SemiAnalysis (opens in a new tab); OfficeChai (opens in a new tab))
The low ceiling is the main takeaway. Labs often promote high scores on older tests, but GDPx shows the strongest models still finish only a small share of long, realistic expert tasks. The cost gap also matters for companies choosing a model, since Gemini 4 Carbon came close to the leader at roughly a third of the price per task.
