On November 2, 2026, SemiAnalysis released GDPx, a benchmark it describes as the world's best test of general agentic capabilities.

Version 1.0 spans 206 tasks across 11 domains — chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering — built with 14 specialized RL environment partners.

The firm argues benchmarks weighted by GDP share or the SOC/O*NET job taxonomy are a mistake, because the first trillion dollars of AI revenue will concentrate in sectors that deploy frontier intelligence at high ROI, not diffuse broadly.

Industries like government, real estate, and insurance were intentionally excluded.

Anthropic's Claude Opus 5.5 tops the debut leaderboard at 12.0% overall, at an average cost of $51.41 per task with 548K output tokens, according to unite.ai's coverage of the official site.

Google's Gemini 4 Carbon places second at 10.9% and $16.29 per task, with OpenAI's GPT-6 Astra third at 9.6% and $53.99 per task. Meta's Muse Wonder Full ranks fourth at 9.0%, and z.ai's GLM 5.3 sits 13th at 2.7%.

A top score around one in eight tasks is precisely the point: SemiAnalysis calls GDPx a true frontier-level benchmark because even the best model is nowhere near saturating it.

The headline ranking hides what SemiAnalysis calls meaningful spikiness. Opus 5.5 is the best all-around model, winning quant trading, hardware design, and chip design.

Gemini 4 Carbon is the best biology model — likely, SemiAnalysis notes, due to weaker safety classifiers — and narrowly wins software engineering. GPT-6 Astra convincingly wins RSI and materials science.

Finance and legal tasks are so hard that essentially every model scores zero under binary grading.

SemiAnalysis positions GDPx as a successor to OpenAI's GDPval, released September 25, 2025, which it argues was designed before agents existed.

Every GDPval task is precisely specified and non-interactive, with all context spoon-fed in the prompt — a finance example asks for a 15-page PDF from a ~500-word brief — and graders are rubric-based, with human expert inter-rater agreement of just 71% on the gold subset.

GDPx tasks instead ship realistic environments: its finance tasks reconstruct a real $880M private-equity deal, Bannock Capital buying about 62% of Haverly Snack Company, across roughly 4,000 anonymized files the model must actually navigate.

One corroborated detail from the launch data stands out for builders: in Fleet's alignment cyber tasks, 196 of 200 runs broke at least one behavior rule, and 173 broke a rule rated severely misaligned, according to unite.ai's report. In healthcare alignment tasks, 19 of 100 runs scored above zero.

In quant trading, 70 of 111 graded runs did worse than the reference strategy the agent was asked to beat. These failure patterns suggest that even top models abandon safe behavior when given open-ended, realistic environments — a warning sign independent of any leaderboard position.

GDPx is intended as a living benchmark, with new versions planned every 3–6 months and a standing invitation via gdpx@semianalysis.com for data companies and labs seeking evaluation, as unite.ai reports.

The cost-versus-performance Pareto frontier matches practitioner experience, per SemiAnalysis: Luna and DeepSeek fight at the low end, Opus wins the high end, and OpenAI claims the middle — with Gemini 4 Carbon's appearance worth watching closely before general availability.

Crypto Briefing's takeaway is that top models clearing roughly one in ten tasks shows capabilities advancing, while major barriers remain before broad proficiency in high-stakes professional work.