GDPx Benchmark Caps Claude Opus 5.5 at Just 12% Overall

The best frontier AI model in the world scores just 12% on real-world economic tasks. That is the headline from GDPx, the agentic benchmark SemiAnalysis released on November 2, 2026. Claude Opus 5.5 leads the overall leaderboard, followed by Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. The benchmark spans 11 domains chosen for their economic impact: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering.

What Makes GDPx Different From Existing Benchmarks

GDPx does not spoon-feed models with prompts. Instead, it builds interactive environments. The finance split drops models into a deal room of roughly 2,000 files from a real private equity transaction representing over $100 billion in enterprise value, including a $30 billion media take-private and an $880 million-plus-earnout acquisition of Haverly Snack Company. Quant trading tasks come via Third Moment, led by Jeffrey Yan, previously Head of Semi-systematic Options at Hudson River Trading. The authors note that when they run their quality assurance process on popular benchmarks like Terminal Bench 4.0, it flags nearly 50% of tasks as broken.

SemiAnalysis assembled GDPx with 14 specialized data partners, including Living Context for finance. The benchmark refreshes every 3 to 6 months to prevent overfitting. Max Kan, Andrew Megalaa, and Dylan Patel authored the release.

Domain-by-Domain Results and Model Strengths

No single model dominates every domain. Claude Opus 5.5 wins quant trading, hardware design, and chip design. Gemini 4 Carbon leads biology and software engineering. GPT-6 Astra tops recursive self-improvement and materials science. The authors report that Opus 5.5 and GPT-6 Astra have sandbagged overall scores due to their safety classifiers, which refuse rather than attempt tasks in cyber, biology, and alignment splits. Muse Wonder Full, with fewer refusals, ranks top on cybersecurity.

Finance and legal returned binary 0% across every model tested. All 13 frontier models failed these domains completely, making them the hardest and most commercially sensitive areas in the benchmark.

Safety Classifiers Distort the Leaderboard

The adjusted leaderboard moves GPT-6 Astra above Claude Opus 5.5 in domains affected by safety refusals. The authors write that Opus 5.5 is the best all-around model when measured on tasks it attempts, but its safety systems prevent it from scoring in certain splits. This highlights a tradeoff: models tuned for safety may underperform on benchmarks that reward task completion over caution.

Release Date and Benchmark Availability

GDPx officially launched on November 2, 2026. SemiAnalysis published the full leaderboard and technical breakdown on the same day. The benchmark is designed for organizations evaluating frontier models for deployment in high-value economic domains. SemiAnalysis plans to refresh the task sets every 3 to 6 months to maintain benchmark validity as models improve.

What This Means for Frontier AI Evaluation

The 12% ceiling exposes how far frontier models remain from human-level performance on economically meaningful tasks. The binary 0% in finance and legal suggests that current models lack the reasoning, domain knowledge, or context handling required for these disciplines. GDPx positions itself as a counterpoint to OpenAI's GDPval, which the authors criticize as non-interactive and context-heavy. The benchmark's environment-based approach may become a new standard for measuring agentic capabilities. Anthropic has been heavily featured in AI news this quarter, with 357 stories in the past 90 days, including recent coverage of Andrej Karpathy shipping HackerTube in roughly two hours with Opus 5.5 just before GDPx landed.