GDPx: frontier agents on high-value domains—and a ~12% ceiling

SemiAnalysis moves the eval story from “average labor in occupational taxonomies” to “where frontier agents can earn high ROI first.” In the public preview, the sharpest number is not a vendor crown—it is the ceiling: top overall score about 12%.

The falsifiable read: if GDPx tasks and grading hold up in the paid text, site, and third-party replication, “general agents nearly deployable in high-value jobs” compresses to “still far below expert workflows.” If 12% is mostly an over-hard ruler, refusal artifacts, or grader noise, that figure alone cannot negate narrow-domain usefulness.

Verified public facts: dated 2 Nov 2026; authors Max Kan, Andrew Megalaa, Dylan Patel; GDPx spans RSI, hardware design, cyber, bio, finance, legal, quant trading, chip design, software engineering, materials science, and alignment. Authors claim the first trillion dollars of AI revenue will concentrate in high-ROI frontier deployments, so they reject GDP-share or SOC/O*NET-weighted “general” benches and intentionally de-emphasize slower-diffusion sectors such as government, real estate, and insurance.

Method preview: partnerships with domain RL-environment startups; partners must pass QA (preferably first try); default ~8-hour limits; native harnesses when available (Claude Code / Codex / Antigravity), else OpenCode. Versus GDPval, they criticize spoon-fed context, non-interactivity, thin environments, rubric noise, bad gold solutions, and broken time-sensitive links.

Evidence class: industry-research newsletter releasing a self-run benchmark—not a peer-reviewed paper. Preview claims include ~12% top overall; OpenAI and Anthropic near the top; SpaceXAI, Meta, and open-weight models meaningfully behind; Gemini 4 Carbon entering the Opus/Astra discussion; cost–performance Pareto with Luna/DeepSeek at the low end, Opus at the high end, OpenAI in the middle.

Limits: the post is marked Paid; this piece uses only publicly visible preview text and chart captions—not the full task table, grading protocol, or raw scores. Model names and domain winners are author assertions pending same-environment third-party replication.

Steelman: if valuable work is long-horizon, multi-tool, and knowledge-base-heavy, GDPval-style “one PDF + one report” systematically overrates looking productive; GDPx-style deal rooms, zero-day labs, and research logs may hit the real deployment bottleneck.

Steelmans’s hard edge is interest and reproducibility: SemiAnalysis sets domain weights, picks partners, runs QA, and sells subscriptions/Tokenomics narrative; safety-classifier “sandbagging” can also decouple overall scores from usable ability. Treat 12% as “unsaturated on their ruler,” not a vendor-neutral physical constant.

For buyers, spikes beat totals: refusals on cyber/bio/alignment can make a “best general model” unusable in the domain that matters. For labs, native-harness scoring binds product wrap to model skill—leaderboard fights will hit toolchains as much as weights.

For policy/safety readers, casting classifiers as an underrated selection variable is both a procurement warning and a potential incentive to loosen refusals for points. What changes the story in six months is independent replication and public task subsets—not another in-house radar chart.

If replications still show frontier models far below expert baselines on GDPx-like long-horizon environment tasks, “high-value job automation at scale” should keep being priced as pilots and human–AI teams—not average GDP displacement. If 12% proves mostly ruler difficulty or refusal artifact, dial down the “universally far from ready” extrapolation—while keeping the methodological claim that rich environments beat one-page prompts.

【No.2】