GDPx is SemiAnalysis’s new benchmark for frontier agentic capability. Version 1.0 spans 206 tasks across 11 domains, built with 14 specialized data partners. On the public leaderboard, Claude Opus 5.5 leads at 12.0% — and SemiAnalysis presents that low ceiling as evidence the test is not yet saturated.
The benchmark launched on November 2, 2026. Primary sources are the public GDPx site and the SemiAnalysis newsletter by Max Kan, Andrew Megalaa, and Dylan Patel. Here is what it measures, how scoring works, and what remains open.
GDPx: Domains and design
SemiAnalysis deliberately does not cover the whole economy. It focuses on domains where the authors expect frontier models to matter most over the next one to two years:
- Chip design
- Cybersecurity
- Biology
- Finance
- Legal
- Quant trading
- Recursive self-improvement (RSI)
- Hardware design
- Materials science
- Alignment
- Software engineering
Government, real estate, and insurance are left out on purpose. The thesis: early AI revenue will concentrate where frontier intelligence has high ROI, not spread evenly by GDP share or SOC/O*NET taxonomies.
Each domain used specialized RL-environment vendors. Named partners include Aria, Arcverex, and Ergodex (chip design), Arithmetic (cyber), Latch Bio (biology), Living Context (finance), Chakra (legal), Third Moment (quant), Extrapolate (RSI), Beautiful Machines (hardware), ValenceForge and Emulated (materials), Fleet (alignment), and Datacurve (software). SemiAnalysis says it approached more than 30 candidates and dropped many after its QA process.
How scoring works
The headline GDPx score averages all 11 domains equally; when a domain has multiple partners, partners are weighted equally too. Each task uses pass@1 across three attempts. Domains mix binary and partial grading. The default time limit is eight hours; some tasks are shorter.
Models ran in native harnesses when available — Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon. Without a native harness, SemiAnalysis defaults to OpenCode. Fireworks provided inference for open-source models; Modal provided sandbox infrastructure.
Safety classifiers can depress overall scores. SemiAnalysis flags Opus 5.5 and GPT-6 Astra in domains such as cyber, biology, and alignment. Excluding those three domains changes rankings. Muse Wonder Full is described as today’s strongest cyber model in the set by a meaningful margin.
Leaderboard: Opus 5.5 leads, far from saturated
The public leaderboard lists overall scores including:
- Claude Opus 5.5 (Anthropic): 12.0%, about $51.41 per task, roughly 548K output tokens
- Gemini 4 Carbon (Google): 10.9%, about $16.29
- GPT-6 Astra (OpenAI): 9.6%, about $53.99
- Muse Wonder Full (Meta): 9.0%, no public price
- Claude Fable 5.1: 8.5%
- GPT 6.1 Sol: 8.4%
- Claude Sonnet 5.5: 8.2%
- Gemini 4 Argon: 7.3%
- DeepSeek 4.1 Flash 1017: 5.4%
- GPT-6 Luna: 5.1%
- Grok 4.7: 3.9%
- GLM 5.3: 2.7%
According to the newsletter, Opus 5.5 wins quant trading, hardware design, and chip design. Gemini 4 Carbon leads biology and edges software engineering. GPT-6 Astra leads RSI and materials science. Finance and legal binary scores are essentially zero; SemiAnalysis also shows partial averages there.
Those ranks align with other public comparisons where Claude Opus 5.5 and GPT-6 sit near the top and cheaper open-weight models such as DeepSeek V4.1-Flash or GLM 5.3 trail. Independent replications of GDPx are not available yet.
Why SemiAnalysis calls GDPval inadequate
Much of the newsletter critiques OpenAI’s GDPval from September 2025. SemiAnalysis calls GDPval impressive for its time but pre-agent: tasks are precisely specified and non-interactive, needed context often sits in the prompt, and the environment stays thin — typically web search, a code interpreter, and document libraries.
GDPx instead ships realistic environments. In Living Context’s finance example, the model must navigate thousands of anonymized deal-room files from a reconstructed PE transaction, rather than one-shot a long PDF from a short prompt. Arithmetic’s cyber tasks are said to use zero-days and multi-service setups. Beautiful Machines’ hardware tasks include tools such as Siemens NX and real AWS FPGA instances. Materials tasks place models in simulated labs.
On QA, SemiAnalysis claims the same process flags nearly half of Terminal Bench 4.0 tasks as broken, and partners had to pass automated QA on the first try with only minor issues. That is a strong vendor claim without independent confirmation.
What the domains actually test
The newsletter and GDPx site describe selected public tasks in more detail. They are illustrative; most tasks stay private.
In RSI, Extrapolate contributes ten tasks. RSI here means an agent’s ability to help build the next version of itself — including debugging failed RL runs, data quality, and inference optimization. Extrapolate stresses that real frontier research usually starts from inherited codebases and logs, not from scratch. One example reconstructs a harness bug that told long rollouts to stop exploring early.
Chip design has three partners. SemiAnalysis says the domain is slightly overrepresented because it expects chip design to be the next major application after coding. Aria, Arcverex, and Ergodex cover verification, PPA optimization, testbenches, and physical layout. Beautiful Machines’ hardware work uses tools such as SOLIDWORKS, Siemens NX, Fusion 360, and KiCad; grading mixes deterministic checks with rubric reviews.
Finance and legal remain the hardest: under binary grading, scores are essentially zero according to SemiAnalysis. Living Context builds 24 finance tasks inside a reconstructed PE transaction with roughly 4,000 files. Chakra supplies legal tasks across eight practice areas with Word and PDF environments. Third Moment’s quant trading scores strategies with a one-sided paired t-test on hidden out-of-sample days.
Fleet’s alignment tasks measure honest, in-bounds behavior under pressure in cyber, insurance, and healthcare worlds. SemiAnalysis argues alignment is not a pure capability metric but still essential for frontier models. Datacurve’s software engineering tasks target real repositories and concise prompts rather than long specifications.
What remains open
GDPx is a proprietary, curated benchmark. Most tasks are private; only selected public tasks are described in depth. SemiAnalysis plans new versions every three to six months. Whether labs will cite GDPx on release cards is unknown.
For readers, the 12% ceiling is not a measure of how much of “the economy” models can do. It measures how often models pass SemiAnalysis’s and partners’ chosen agent tasks on pass@1 across three runs. Domain selection and safety classifiers move the ranks. Treat GDPx as one more data point — alongside your own tests and established boards such as our LLM leaderboard.
