SemiAnalysis launches GDPx, and the best models top out at 12%

On November 2, SemiAnalysis released GDPx, a benchmark built to measure how well frontier models handle agentic work in real professional settings. The debut leaderboard opened with an uncomfortable number: the top-scoring model completed just 12% of its tasks.

The first version is substantial. GDPx spans 11 professional domains, 206 tasks contributed with 14 specialised data partners, and a leaderboard of 13 models. Unlike question-and-answer evaluations, tasks carry a default eight-hour window, and models run in their native harness where one exists — Claude Code, Codex, Antigravity — with OpenCode as the fallback, so they must plan, use tools and hand over a deliverable rather than answer a single prompt.

The ranking itself tells the story. Anthropic's Claude Opus 5.5 leads at 12.0%, Google's Gemini 4 Carbon follows at 10.9%, and OpenAI's GPT-6 Astra sits third at 9.6%, while Z.ai's GLM 5.3 anchors the table at 2.7%. Even the strongest systems available today fail to produce acceptable deliverables on most tasks.

Domain-level results are harsher than the average. In finance and legal, every model scores zero under the intended binary grading, forcing SemiAnalysis to report partial scores instead. In quant trading, 70 of 111 graded runs did worse than the reference strategy the agent was asked to beat. These are not near misses; they are failures on some of the most valuable work on offer.

The methodology carries the ambition. The authors position GDPx as a successor to OpenAI's earlier GDPval, arguing that GDPval spoon-feeds context into the prompt and never forces real interaction. GDPx instead tries to rebuild the job site: its 24 finance tasks reconstruct a fund's purchase of roughly 62% of a snack manufacturer for about $880 million plus an earnout, spread across some 4,000 anonymised files, and each task is scored by its intended method, binary or continuous.

The point of a 12% ceiling is not who tops the table but how far the industry still is from handing real professional work to agents. For buyers, it means agents remain assistants in high-stakes finance and legal work, not executors. It also hands vendors an unavoidable reference point: SemiAnalysis notes that the totals for Opus 5.5 and GPT-6 Astra are dragged down by their own safety classifiers on cyber, biology and alignment tasks, which quietly makes safety policy part of the score.

Three things to watch next. Whether independent teams can reproduce these numbers; whether the promised refresh every three to six months keeps raising the bar on long-horizon, environment-heavy work; and whether labs begin citing GDPx on launch cards. If they do, the benchmark will have bought itself real authority.