No AI model scored above 12% on SemiAnalysis’s real-work benchmark
Reported by SemiAnalysis, FourWeekMBA, news.lavx.hu and 4 more outlets7 sources ↓
SemiAnalysis’s GDPx benchmark put 13 AI models through 206 simulated professional tasks across 11 domains. Claude Opus 5.5 led with 12%; no model fully completed the finance or legal tasks.
Key points
- Gemini 4 Carbon scored 10.9% and GPT-6 Astra 9.6%, according to The Decoder
- Gemini 4 Carbon cost $16.29 per task, about a third of Opus 5.5’s cost, The Decoder reports
- GDPx tasks use simulated work environments and allow up to eight hours
- SemiAnalysis says it will update the benchmark every three to six months
Sources
- SemiAnalysisRead ↗GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains
- FourWeekMBARead ↗Opus 5.5 Leads SemiAnalysis’s GDPx Agent Benchmark at 12%
- news.lavx.huRead ↗GDPx Benchmark Measures Frontier AI Across Most Valuable Domains
- ShiftDelete.NetRead ↗GDPx AI Benchmark Tops Out at 12% as Claude Opus 5.5 Leads
- SSBCrackRead ↗GDPx Benchmark Targets Frontier Intelligence in Key Domains
Details
GDPx evaluates models in simulated work environments with professional tools, files and other task-specific context, rather than giving them only a prompt. Tasks cover 11 areas, including chip design, cybersecurity, biology, finance, law and software engineering. SemiAnalysis developed the benchmark with 14 specialized partners.
The Decoder reports that finance and legal tasks received partial-credit scoring because no model passed them under the benchmark’s strict binary standard. SemiAnalysis says the top three overall scores were close, and that removing cyber, biology and alignment tasks changes the ranking: GPT-6 Astra scores 11.8% in that filtered analysis.
The Decoder notes questions about the benchmark’s comparability and verification: models use their providers’ own agent software, some domains have about ten tasks, and some results are graded by language models. The tasks are mostly not public, which limits outside scrutiny but also makes it harder for labs to optimize specifically for them.
Written from coverage as of
Images
- The Decoder
- Tech Times
- SemiAnalysis
- The Decoder
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh