SemiAnalysis's New GDPx Benchmark: The Best Frontier Model Scores Just 12%
The 206-task agentic benchmark spans 11 high-value fields, from chip design to quant trading. Claude Opus 5.5 leads, and finance and legal tasks score zero under strict grading.
Published
Last updated / Fact-checked on
SemiAnalysis has launched GDPx, a benchmark of 206 realistic agentic tasks across 11 fields it expects to drive the most AI revenue. The top model, Anthropic's Claude Opus 5.5, scores 12.0%, and almost every model scores zero on finance and legal tasks under the intended strict grading.
Photo: Akela999 / Pixabay
The leaderboard
The GDPx site ranks 20 models. These are the top eight:
| Rank | Model | GDPx score |
|---|---|---|
| 1 | Claude Opus 5.5 | 12.0 |
| 2 | Gemini 4 Carbon | 10.9 |
| 3 | GPT-6 Astra | 9.6 |
| 4 | Muse Wonder Full | 9.0 |
| 5 | Claude Fable 5.1 | 8.5 |
| 6 | GPT 6.1 Sol | 8.4 |
| 7 | Claude Sonnet 5.5 | 8.2 |
| 8 | Gemini 4 Argon | 7.3 |
Further down, GPT-6 Luna scores 5.1, Grok 4.7 3.9 and GLM 5.3 2.7. "GDPx is a true frontier-level benchmark," SemiAnalysis wrote in its November 2 newsletter.
What GDPx measures
SemiAnalysis argues that "the first trillion dollars of AI revenue will be concentrated in sectors that can deploy frontier intelligence at high ROI." That means benchmarks built on share of GDP or standard occupational lists miss the point. GDPx covers chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment and software engineering. It deliberately skips government, real estate and insurance.
Every task sits inside a working environment instead of a self-contained prompt. Finance tasks put the model inside a replica private-equity deal, with about 4,000 anonymized files seeded from a real transaction. Cyber tasks are built on zero-day vulnerabilities found by the partner Arithmetic. SemiAnalysis contrasts this with OpenAI's GDPval, released in September 2025, whose tasks it says are "precisely specified and non-interactive" and hand the model all the context it needs.
How it was built and scored
The tasks come from 14 specialized data partners:
- Chip design: Aria, Arcverex and Ergodex
- Cybersecurity: Arithmetic
- Biology: LatchBio
- Finance: Living Context
- Legal: Chakra
- Quant trading: Third Moment
- RSI: Extrapolate
- Hardware design: Beautiful Machines
- Materials science: ValenceForge and Emulated
- Alignment: Fleet
- Software engineering: Datacurve
SemiAnalysis says it contacted more than 30 companies. Partners had to pass its automated quality checks on the first try.
The default time limit is eight hours per task. Models ran in their native harness where one exists, such as Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon. The headline score is pass@1 across three attempts. All 11 domains count equally, and partners within a domain count equally.
Where models shine and fail
SemiAnalysis calls Opus 5.5 "the best all-around model," winning quant trading, hardware design and chip design. Gemini 4 Carbon is the best biology model, which SemiAnalysis says is likely because of weaker safety classifiers, and narrowly wins software engineering. GPT-6 Astra "wins convincingly" on RSI and leads materials science.
Finance and legal tasks are so hard that "basically every model scores exactly 0" under binary grading, so the leaderboard shows average partial scores for those two domains.
SemiAnalysis also says safety classifiers drag down overall scores for Opus 5.5 and GPT-6 Astra in cyber, biology and alignment. It names Meta's Muse Wonder Full as "the best available cyber model today by a meaningful margin."
How to use these results
- Pick models by domain, not by the headline number. Scores vary widely across fields.
- Check safety-classifier effects if you work in security or biology, where refusals can hide real capability.
- Don't trust agents on high-stakes finance or legal work without supervision. No model passed those tasks under strict grading.
- Remember the results are self-reported. SemiAnalysis designed, ran and published the benchmark. It plans new versions every three to six months.
For another example of a company grading models on its own finance benchmark, see our report on Samaya Sterling and FrontierFinance.
More coverage is on our Artificial Intelligence hub.
How we reported this
We read SemiAnalysis's full November 2 newsletter and the live GDPx leaderboard, and we took every score from the leaderboard as of publication. We corrected the scoring description to match SemiAnalysis's methodology: pass@1 across three attempts, with domains weighted equally. No one has independently replicated these results.
Published and fact-checked November 2, 2026.
Sources
- SemiAnalysis, Max Kan, Andrew Megalaa and Dylan Patel, "GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains," Nov. 2, 2026: https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities
- GDPx official leaderboard: https://gdpx.semianalysis.com/
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.