SemiAnalysis has launched the GDPx benchmark, a new test that measures how well frontier AI models handle long, economically valuable agentic work. Published November 2, 2026, the first results show Anthropic’s Claude Opus 5.5 as the best all-around model, but even the leader scores only about 12%.
What Is the GDPx Benchmark?
GDPx covers 11 domains chosen for their economic impact: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment and software engineering. SemiAnalysis argues AI revenue will concentrate where frontier models deliver the highest returns, so it picked domains by impact rather than by share of GDP.
Tasks are realistic and interactive. One finance task, for example, has models navigate a simulated deal room with roughly 4,000 files. Each task has a default eight-hour limit, and models run in their own native tools where possible.
GDPx Results
- Claude Opus 5.5 is the best overall model, at about 12.0%.
- Gemini 4 Carbon follows at about 10.9% and leads in biology, narrowly edging out others in software engineering.
- GPT-6 Astra scores about 9.6% and leads in RSI and materials science.
- Muse Wonder Full is described as the best cybersecurity model by a clear margin.
SemiAnalysis notes that Opus 5.5 and GPT-6 Astra lose points because their safety classifiers interfere with some cyber, biology and alignment tasks. Finance and legal tasks proved so hard under pass/fail grading that nearly every model scored zero, so partial scores are reported for those.
Cost vs. Performance
On value for money, lower-cost models from Luna and DeepSeek lead at the budget end, Opus leads at the high end, and OpenAI’s models sit in the middle. That lines up with the ongoing price war on cheaper AI models and launches like GPT-6.1 Sol.
A Swipe at GDPval
The authors argue that OpenAI’s GDPval benchmark is outdated because it is single-shot, gives models all context up front and relies on noisy rubrics, citing 71% human agreement on its gold subset. GDPx instead makes models work through realistic environments.
Notable Failures
- In biology, models often made statistical errors and missed instrument artifacts.
- In hardware design, some models struggled with assembly interfaces, while Opus and Astra produced designs that could be rebuilt.
- In RSI tasks, models often failed to connect a harness bug to degraded training.
Why It Matters
A 12% top score is a reminder that today’s AI models still struggle with long, real-world professional tasks. For SEO and marketing teams, it’s a useful check on expectations: AI is strong at drafting and analysis, but complex multi-step work still needs human review. Accuracy problems show up in everyday use too, as a new study of chatbot election answers found.
Key Takeaways
- GDPx tests frontier models on 11 economically valuable domains.
- Claude Opus 5.5 leads at about 12%, followed by Gemini 4 Carbon and GPT-6 Astra.
- Finance and legal tasks scored near zero across models.
Source: SemiAnalysis.
One Response