SemiAnalysis launches GDPx: the best frontier AI model scores just 12% across 11 real-world domains

GDPx covers 11 domains SemiAnalysis picked for where frontier intelligence will earn its highest return, rather than for their share of GDP: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Government, real estate and insurance are deliberately left out. SemiAnalysis argues the first trillion dollars of AI revenue will concentrate in sectors that can deploy frontier intelligence at high return on investment, not diffuse evenly across the economy.

That framing is a direct challenge to OpenAI’s GDPval, the benchmark SemiAnalysis says remains the most widely reported general-purpose yardstick despite being released on September 25, 2025, before agents were central to frontier models. GDPval tasks, SemiAnalysis argues, spoon-feed the model all the context in the prompt and ask for a one-shot deliverable, with no simulated work environment to explore. GDPx tasks, by contrast, put the agent inside a detailed environment and require it to find the relevant information and act.

The leaderboard: spiky, not settled#

On GDPx v1.0, Anthropic’s Claude Opus 5.5 leads overall at 12.0%, ahead of Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%. No model dominates: Opus 5.5 wins quant trading, hardware design and chip design; Gemini 4 Carbon leads biology and edges out software engineering; GPT-6 Astra wins recursive self-improvement convincingly and leads materials science. SemiAnalysis calls the pattern “meaningful spikiness” across all leading models.

The zeros are as telling as the wins. SemiAnalysis says its finance and legal tasks are so challenging that basically every model scores exactly 0. And safety classifiers distort the picture: Opus 5.5 and GPT-6 Astra have “sandbagged” overall scores because their safety classifiers sometimes refuse in cybersecurity, biology and alignment, SemiAnalysis says. Exclude those three domains and the ranking shifts. The practical consequence, the firm argues, is that model choice depends on the use case — Meta’s Muse Wonder Full, with fewer refusals, is “actually the best available cyber model today by a meaningful margin.”

Cost tells a similar story. On the cost-versus-performance frontier, SemiAnalysis says Luna and DeepSeek “fight for the low end,” Opus wins the high end and OpenAI claims the middle — while Gemini 4 Carbon is “a real improvement from Google and the only model in the same ballpark as Opus and Astra.” SemiAnalysis adds that it will be interesting to see people’s impressions once Carbon becomes generally available.

Built with specialists, checked hard#

SemiAnalysis did not build the tasks alone. It partnered with 14 specialist startups that build reinforcement-learning environments — three for chip design alone (Aria, Arcverex and Ergodex), plus Arithmetic for cybersecurity, Latch Bio for biology, Living Context for finance, Chakra for legal, Extrapolate for recursive self-improvement, Beautiful Machines for hardware design, ValenceForge and Emulated for materials science, Fleet for alignment and Datacurve for software engineering. Each partner had to pass SemiAnalysis’s automated quality checks on the first try, with only minor mistakes at worst; partners that failed, including some prominent ones, were dropped.

The firm says the same quality process flags nearly 50% of tasks in Terminal Bench 4.0 as broken — its argument for why a new benchmark was needed, and why task quality, not just task count, decides whether a leaderboard means anything.

The finance tasks show what “realistic” means here. One example, contributed by Living Context, is seeded by a real transaction: Bannock Capital, a private-equity fund, buys about 62% of Haverly Snack Company for $880 million plus an earnout. The environment holds roughly 4,000 anonymised files across the target’s virtual data room and separate deal folders for each party, and the agent must navigate a roughly 2,000-file deal room to gather context — the way a human associate on the real deal did — before completing a task that associate was actually asked to do.

Tasks run with a default time limit of 8 hours, in each model’s native harness where one exists — Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon — defaulting to OpenCode otherwise. Grading mixes binary and partial credit depending on the domain, but the headline score is pass@1 across three attempts, with all domains weighted equally.

What 12% actually means#

Twelve percent is not a claim that frontier models are useless at professional work. It is a claim about completing multi-step work end to end, inside a rich simulated environment, against a task verifier — the kind of work where finding the right context is half the job. On that measure, the best agent in the world finishes roughly one task in eight, and the gap between the best and the rest is a few percentage points, not a chasm.

SemiAnalysis plans to update GDPx every 3 to 6 months, incorporating feedback and adding partners. If the benchmark holds up, that cadence — not any single leaderboard — is what will show whether agents are closing the gap between answering well and working well. For now, the message for anyone buying frontier models for real work is narrower than the marketing: pick by domain, test in your own environment, and do not mistake a quiz score for a finished job.

Sources: GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains (SemiAnalysis), November 2, 2026; SemiAnalysis unveils GDPx benchmark, and the best AI model scores 12% (Crypto Briefing), November 2, 2026; SemiAnalysis Launches GDPx; Best Frontier Model Hits 12% (AI Weekly), November 2, 2026 — checked November 2, 2026.