GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

What it claims

GDPx is a SemiAnalysis benchmark of long-horizon agentic tasks. The best model scores 12%. SemiAnalysis argues that the first trillion dollars of AI revenue will concentrate in a few high-ROI sectors, so it chose domains on that basis rather than by share of GDP or the SOC/O*NET taxonomy. It deliberately leaves out government, real estate and insurance. The headline ranking is distorted by safety classifiers: Opus 5.5 and GPT-6 Astra lose points in cyber, biology and alignment, so the article also gives a ranking with those three domains excluded.

Per-domain winners, as stated:

• Opus 5.5: best all-around model. Wins quant trading, hardware design and chip design.
• GPT-6 Astra: wins recursive self-improvement (RSI) "convincingly" and materials science.
• Gemini 4 Carbon: wins biology (attributed to weaker safety classifiers) and narrowly wins software engineering.
• Muse Wonder Full: best cyber model "by a meaningful margin."
• Finance and legal: nearly every model scores exactly 0.

On cost versus performance, Luna and DeepSeek compete at the low end, Opus leads the high end, and OpenAI holds the middle.

How it works

Task sourcing and QA. SemiAnalysis approached more than 30 RL-environment vendors and ran every one through an automated agentic QA process. The process checks for consistency, completeness, cheating paths and unfair verifiers. Partners had to pass on the first try with only minor errors. The reasoning is that a task which passes only after several rounds of feedback may have been tuned to the checker rather than fixed. Prominent vendors were dropped under this rule. The same QA flagged nearly 50% of Terminal Bench 4.0 tasks as broken. Hundreds of trajectories were also reviewed by hand.

Run settings.

• The default time limit is 8 hours per task.
• Each model runs in its native harness: Claude Code, Codex or Antigravity. Models without one run in OpenCode.
• The headline score is pass@1 across three attempts.
• Domains are weighted equally. Where a domain has several partners, each partner is weighted equally.
• Chip design has three partners, so it is slightly overrepresented, and the article says this is intentional.

Environments.

• Finance: 24 tasks set inside a reconstructed private-equity deal of about 4,000 files. Files are dated to what the analyst would have known on the task date.
• Cyber: built on Arithmetic's private zero-days, running in multi-service enterprise environments. Only about 5% of the zero-days Arithmetic finds become tasks.
• Hardware: live SOLIDWORKS, Siemens NX, Fusion 360, KiCad, Vivado and Cadence sessions.
• Biology: 13 tasks graded deterministically against real wet-lab data, deliberately without rubric judges.

Evidence

Results on individual public tasks:

• Ergodex vt01, a timing ECO on an Ibex RISC-V core in Nangate45 with an 8-hour, 8-CPU budget and at most 104 cells changed. GPT-6 Astra averaged 0.34, Claude Opus 0.135 and Fable 0.07. Ergodex says OpenAI models consistently beat Anthropic models on ECO tasks.
• Living Context Bellwood-QoE, 25 attempts with no solves. Averages out of 11 checks: GPT 6.1 Sol 5.6, Opus 5.5 and GPT-6 Astra 5.4, GLM 5.3 and Grok 4.7 under 2. The best single run passed 8.
• Beautiful Machines compressor housing, 12 runs with no verified full pass. Five of six Opus and Astra submissions passed both the mechanism-clearance checks and every edit-perturbation check.
• Extrapolate RSI GRPO-debugging task, 20 rollouts across four models. All four models struggled.
• Arithmetic time-to-exploit: GPT-5.5 needed more than three hours. GPT-5.6 Luna and Sol took 35 to 50 minutes.

Limits

• Safety classifiers: headline scores mix capability with refusals. Choose a model by per-domain score, not by overall rank.
• Finance scoring: finance is nominally binary-graded, but every model scores 0 that way, so the leaderboard reports average partial score instead.
• Grading: grading is mixed, binary or partial depending on the domain, and hardware uses LLM judges alongside deterministic checks. This is the same rubric-noise concern the article raises against GDPval.
• Small task counts: RSI has 10 tasks and biology 13.
• Partner-written sections: the domain sections were "largely written by the partners." Those partners sell data to labs.
• Recorded failure modes:
• Finance: models trust the most explicit figure over the authoritative document.
• Biology: pseudo-replication in statistics.
• Chip design: weak verification under PPA pressure.
• RSI: models do not link a harness message to agent behaviour.
• Missing data: full numeric leaderboards appear only as images or on the official site, not in the text.
• Truncated page: the fetched page covers about 60,000 of roughly 90,000 characters. Materials science is cut off, and the alignment, legal, quant trading and software engineering sections are missing.