The Pulse
SemiAnalysis GDPx Gives Its Top AI Agent a 12% Score
Claude Opus 5.5 leads SemiAnalysis’ GDPx benchmark with a 12% aggregate score across 11 domains. The benchmark uses different grading methods, and its finance and legal results include partial scores alongside zero scores under intended bin
AI.info Team · Written with AI · fact-checked ·
Claude Opus 5.5 scores 12% on GDPx, the highest result in SemiAnalysis’ new benchmark of AI agents working across 11 specialized domains. The score is an aggregate benchmark result, not a claim that the model completes exactly 12% of all professional work: GDPx combines tasks with different grading methods and weights each domain equally.
SemiAnalysis published the benchmark on Nov. 2, 2026, with a leaderboard covering 20 models and a public dashboard listing 206 tasks contributed by 14 specialist partners. Gemini 4 Carbon scores 10.9%, GPT-6 Astra 9.6% and Muse Wonder Full 9.0%, behind Opus. The results offer a demanding snapshot of agent performance on complex work, but the benchmark’s design and caveats matter when comparing the scores.
GDPx tests work across 11 specialized domains
The benchmark spans chip design, cybersecurity, biology, finance, legal work, quantitative trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis says it chose domains where it expects frontier AI to have high economic value, rather than building a broad test based on each occupation’s share of GDP. The newsletter says the benchmark forgoes government, real estate and insurance as industries, while its alignment section separately describes three insurance tasks.
Tasks come from 14 specialist partners, including LatchBio for biology, Third Moment for quant trading, and Living Context for finance. Some environments put agents inside simulated laboratories or engineering tools; a finance example draws on thousands of anonymized transaction files. The default time limit for a GDPx task is eight hours, though some tasks have shorter limits. Models run in their native agent harness when available; otherwise, SemiAnalysis says it defaults to OpenCode.
GDPx is a true frontier-level benchmark: even the best model scores just 12%.
Max Kan, Andrew Megalaa and Dylan Patel, SemiAnalysis authors
One overall score hides sharp differences
The top model varies by domain. SemiAnalysis says Opus 5.5 leads in quant trading, hardware design and chip design; GPT-6 Astra leads in recursive self-improvement and materials science; and Gemini 4 Carbon tops biology and narrowly leads software engineering. The newsletter says finance and legal tasks are too difficult for all models under the intended binary grading, while reporting average partial scores for both domains.
Scoring is not uniform. The headline uses pass@1 across three attempts, with all 11 domains weighted equally, while individual tasks use either binary or partial grading. In finance, SemiAnalysis says strict pass-or-fail scoring would give every model zero, so it reports average partial scores instead. It describes the legal tasks similarly: all models are too weak under intended binary grading, and the leaderboard also reports average partial scores. A 12% aggregate therefore cannot be read as a simple percentage of tasks completed successfully.
Safety filters complicate the leaderboard
The authors say safety classifiers lower the overall scores of Opus 5.5 and GPT-6 Astra in cybersecurity, biology and alignment. Removing those three domains changes the comparison, though the article does not state exact adjusted scores in its text. SemiAnalysis also points to “spikiness” in model performance: a high overall rank does not mean a model leads in every kind of work.
Those caveats sit alongside a methodological choice that affects comparisons: models ran in different native harnesses where available. SemiAnalysis also says its tasks underwent automated quality checks and manual review, but GDPx is a new benchmark built by the company and its partners, not an independent assessment of the benchmark’s own grading choices. The public dashboard provides the domain-level results for readers to inspect alongside the headline ranking.
The benchmark is designed to change
SemiAnalysis describes GDPx as a living benchmark and says it plans updates every three to six months as model capabilities change. That cadence will make later results useful only if readers can track what changed between versions: the tasks, partner contributions, model harnesses and scoring rules. For this first release, the clearest finding is the gap between the top aggregate score and a perfect result: Opus 5.5 earns 12% across the benchmark’s 11 domains.