GDPx, SemiAnalysis's New Benchmark: Even Top-Ranked Claude Opus 5.5 Scores Just 12%
On November 2, SemiAnalysis released GDPx, a benchmark of 206 agentic tasks spanning 11 high-value domains (chip design, cybersecurity, biology, finance, law, etc.). Claude Opus 5.5 took the top spot with 12.0, ahead of Gemini 4 Carbon (10.9) and GPT-6 Astra (9.6). Built-in safety filters weighed heavily on the rankings.
Boutique research firm SemiAnalysis released GDPx on November 2, 2026, an AI agent benchmark designed as the successor to OpenAI’s GDPval, which launched in September 2025. The verdict from version 1.0: no model scored above 12%. Claude Opus 5.5 took first place with 12.0, ahead of Gemini 4 Carbon (10.9) and GPT-6 Astra (9.6).
206 Tasks in the Domains Expected to Generate Real ROI
GDPx features 206 tasks across 11 domains selected because, according to SemiAnalysis, the earliest massive revenue streams for AI will emerge from sectors that can deploy frontier AI with strong returns on investment: chip design, cybersecurity, biology, finance, law, quantitative trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. Administrative workflows, real estate, and insurance were deliberately left out.
The critical shift from GDPval lies in the environments. Instead of detailed text prompts with attached files, each task drops the agent into a simulated workspace: an M&A data room holding roughly 4,000 files in finance, live computing infrastructures running unpatched zero-day exploits in cybersecurity, or active CAD software sessions in hardware design. The tasks were supplied by 14 specialized firms that build training environments for frontier labs. Each model gets up to 8 hours per task, executing within its native environment where one exists (Claude Code, Codex, Antigravity). The headline metric is pass@1 averaged over three attempts, with equal weight across all domains.
Highly Divergent Strengths Across Domains
| Model | GDPx Score |
|---|---|
| Claude Opus 5.5 | 12.0 |
| Gemini 4 Carbon | 10.9 |
| GPT-6 Astra | 9.6 |
| Muse Wonder Full | 9.0 |
| Claude Fable 5.1 | 8.5 |
| GPT 6.1 Sol | 8.4 |
| Claude Sonnet 5.5 | 8.2 |
| Gemini 4 Argon | 7.3 |
| Grok 4.7 | 3.9 |
| GLM 5.3 | 2.7 |
Excerpt from the overall leaderboard on the GDPx website, accessed November 2, 2026.
Behind the overall averages lie starkly specialized profiles. According to SemiAnalysis, Opus 5.5 won in quantitative trading, hardware design, and chip design; Gemini 4 Carbon led in biology and, narrowly, in software engineering; GPT-6 Astra came out on top in recursive self-improvement and materials science. In finance and law, tasks proved so demanding that nearly every model scored zero under the planned binary grading, forcing the leaderboard to rely on partial credit for those two categories.
Safety Filters Skew the Results
SemiAnalysis emphasized that scores for Opus 5.5 and GPT-6 Astra were dragged down by their safety classifiers, which frequently tripped up execution in cybersecurity, biology, and alignment. As a result, the firm published a secondary leaderboard excluding those three domains. It also noted that Muse Wonder Full is currently the top-performing available model in cybersecurity, while attributing Gemini 4 Carbon’s win in biology likely to less stringent guardrails. Gemini 4 Carbon, announced in late October with initially restricted access, is not yet broadly available.
Our take
GDPx succeeds at measuring what enterprises actually want to buy: an agent capable of hunting down information across realistic environments before delivering a finished product. Scores hovering around 10% serve as a sharp reminder that full autonomy across multi-hour, expert-level tasks remains a distant target. Two caveats apply: the benchmarks are largely private and sourced from firms that sell data to the labs, and the findings have not been independently reproduced. The most actionable takeaway for enterprise buyers lies elsewhere: with safety guardrails penalizing scores in critical sectors, choosing the “best model” depends just as much on a developer’s moderation policy as on raw architectural capability.
Sources
- GDPx — Official Website and Leaderboard (SemiAnalysis) Primary source
- SemiAnalysis — GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains (November 2, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
