Twelve percent is an average of eleven different tests
The headline number is real, but it needs context. GDPx gives each of its 11 domains equal weight and reports pass@1 across three attempts. Finance and legal are particularly difficult: SemiAnalysis says every model scores zero under the intended binary grading in both domains. It therefore also reports average partial scores for those areas. The headline is consequently a composite of domains with different grading approaches, not a simple count of tasks fully completed. SemiAnalysis chose these domains because it believes they are where frontier AI could generate the most economic value, not because they represent the economy evenly.
One finance task drops the agent into a deal room of roughly 2,000 files from a reconstructed real transaction and asks for work an associate actually did. That is the distinction SemiAnalysis draws against OpenAI's GDPval, released in 2025: its criticism is that GDPval gives the model a precisely specified, non-interactive task with the necessary context supplied in the prompt, rather than requiring it to navigate a working environment and gather evidence itself.
So read 12.0% as a statement about the distance between frontier agents and unsupervised professional work in the hardest settings, not as a general ability score. On the domains where work is more tractable, the picture is spiky rather than uniformly bleak: Opus 5.5 leads in quant trading, hardware design and chip design, Gemini 4 Carbon in biology and software engineering, GPT-6 Astra in recursive self-improvement and materials science. No model leads everywhere, which is itself information a single average hides.
Two leaderboards, and the difference is refusal
SemiAnalysis flags its own caveat in the announcement: Opus 5.5 and GPT-6 Astra, it says, have "sandbagged" overall scores because their safety classifiers cut in on cybersecurity, biology and alignment tasks. The same effect explains the strangest line in the release: Muse Wonder Full, fourth overall at 9.0%, is named the best available cybersecurity model by a meaningful margin.
SemiAnalysis also publishes a second ranking that excludes cybersecurity, biology and alignment, the three domains where it says safety classifiers can intervene. The order changes when those domains are removed. This is a useful alternative view, but it is not a definitive measure of underlying capability: removing the domains also removes work that some real customers may need models to perform. The two rankings answer different questions — how models perform under their actual safety behavior, and how they compare on the remaining domains.
Buyers should notice what this means in practice. Part of what a GDPx score measures is not whether a model can do a task but whether its maker will let it try. For a lab choosing a model for security or biology work, the adjusted table is the relevant one, and the unadjusted headline penalizes exactly the models with the strictest refusal behavior. Neither ranking is the honest one on its own; they answer different questions.
Who built the tasks matters too
The tasks were built with 14 specialist data and RL-environment partners. Some operate in the same commercial market as the labs being evaluated: SemiAnalysis describes partners supplying training data, evaluation environments and other data services to frontier AI companies. The firm also sells research subscriptions and has a commercial interest in making GDPx a benchmark that customers follow. These relationships do not establish that any model received preferential treatment, but they are relevant context when assessing a benchmark whose task suppliers also participate in the frontier-AI data market.
The firm did apply a real quality bar, requiring partners to pass its automated QA on the first attempt and dropping prominent vendors that failed, and it notes labs pay $5,000 or more for a single task of this quality. That is a stronger process than most benchmark launches describe.
SemiAnalysis says it plans to release new versions every three to six months. That cadence will provide a useful test of whether the benchmark can track progress without making comparisons across versions meaningless.
Sources
- [1] SemiAnalysis newsletter — "GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains," November 2, 2026 (domains and exclusions; 12% headline; safety-classifier caveat and second ranking; domain winners; finance and legal zero under binary grading, with average partial scores also reported; Muse Wonder Full as best cyber model; methodology — 8-hour default limit, native harnesses, pass@1 across three attempts, equal domain weighting; partner QA first-try rule; $5k+ per task; finance deal-room environment; planned 3-6 month refresh; read first-hand)Read source
- [2] GDPx official site (v1.0; 11 domains, 14 partners, 206 tasks; domain-comparison table lists 13 models, with 20 in the cost-performance view; Muse Wonder Full not shown on the cost chart, no public price; read first-hand)Read source
- [3] Unite.AI — "SemiAnalysis Launches GDPx Benchmark Spanning 11 Agentic Domains," November 2, 2026 (leaderboard as published on the official site: Muse Wonder Full 4th at 9.0%, GLM 5.3 13th at 2.7%)Read source