Claude Opus 5.5 leads GDPx v1.0 with a 12.0% score, ahead of Gemini 4 Carbon at 10.9% and GPT-6 Astra at 9.6%. SemiAnalysis introduced the benchmark on November 2, 2026, bringing together 206 tasks across 11 domains and 14 specialized data partners. The results give Anthropic the overall lead, but leave substantial room for improvement across every competing model. (GDPx leaderboard)
The more useful story sits underneath that ranking. GDPx targets difficult, economically valuable agent work, including chip design, cybersecurity, biology, and quantitative trading. Its published results show models leading in different specialties, safety classifiers affecting performance, and finance and legal tasks producing almost no success.
A 12% score on this deliberately difficult collection deserves attention. It also needs careful interpretation: GDPx measures performance on its selected tasks and environments, not the percentage of the economy that AI can automate.
What GDPx Actually Measures
SemiAnalysis built GDPx around a specific economic thesis: frontier AI will generate its earliest large revenues in sectors where advanced intelligence can deliver especially high returns. That is the authors’ rationale for selecting these domains, not an independently established forecast of where AI revenue will land. (SemiAnalysis announcement)
GDPx covers:
- Engineering: chip design, hardware design, software engineering, and materials science.
- Scientific and security work: biology and cybersecurity.
- Commercial specialties: finance, legal work, and quantitative trading.
- AI development: alignment and recursive self-improvement, or RSI.
The distinction from a short-answer benchmark matters. An environment-based task can require an agent to locate relevant information, use tools, make changes, and produce an outcome that survives grading. Knowing an answer and completing a workflow are different capabilities.
SemiAnalysis positions GDPx as a successor to the kind of broad knowledge-work evaluation represented by OpenAI’s GDPval. Its emphasis is on general agentic capability within demanding specialties, with tasks supplied by organizations that know those specialties.
The 14 partners include Arithmetic for cybersecurity, LatchBio for biology, Chakra for legal work, Living Context for finance, and Third Moment for quantitative trading. Multiple partners contributed to chip design and materials science. (GDPx)
That structure is sensible in my view. A convincing chip-design evaluation needs more than a difficult prompt: it needs a realistic problem, a usable environment, and a grader that recognizes valid solutions. Domain expertise matters at all three stages.
What the 12% Lead Means
The published overall standings put three models at the front:
| Model | GDPx overall score |
|---|---|
| Claude Opus 5.5 | 12.0% |
| Gemini 4 Carbon | 10.9% |
| GPT-6 Astra | 9.6% |
Source: GDPx’s published leaderboard. These are the benchmark publisher’s results, not a separate evaluation conducted for this article.
Opus leads Carbon by 1.1 percentage points and Astra by 2.4 percentage points. Those gaps establish the ordering on the published leaderboard. They do not, by themselves, tell us how reliably the same ordering would hold under different tasks, environments, or grading choices.
The low ceiling is useful because it leaves room to measure improvement. A benchmark that most leading systems nearly perfect has limited ability to distinguish their remaining weaknesses. GDPx instead presents a set of problems that current systems frequently fail to complete successfully.
Low scores alone, however, cannot validate a benchmark. SemiAnalysis explicitly discusses the danger of unfair grading: a task might allow two legitimate solutions while its grader accepts only one. That would measure conformity to a hidden requirement instead of actual competence. The authors say their partner tasks underwent a quality-assurance process intended to address such problems. (SemiAnalysis)
The public materials available here do not expose every task and rubric sufficiently for full independent reconstruction. That limits how confidently an outside reader can separate genuine difficulty from evaluation-design effects.
My reading is that GDPx provides evidence of a substantial remaining gap on this particular set of specialist workflows. Turning that into a claim that Claude can automate exactly 12% of valuable professional work would go well beyond the evidence.
Domain Winners Change the Buying Decision
The overall winner is not the winner everywhere. SemiAnalysis describes pronounced differences between models’ strengths:
| Model | Leading domains reported by SemiAnalysis |
|---|---|
| Claude Opus 5.5 | Quantitative trading, hardware design, chip design |
| Gemini 4 Carbon | Biology, software engineering |
| GPT-6 Astra | Recursive self-improvement, materials science |
| Muse Wonder Full | Cybersecurity |
Domain leadership is based on SemiAnalysis’s published analysis.
These peaks make the aggregate leaderboard less useful as a universal purchasing guide. A chip-design team and a materials-science team could reasonably start their evaluations with different models, even though Opus leads overall.
That is the practical implication I would prioritize. Use the relevant domain result to form a shortlist, then evaluate the actual workflow, tools, permissions, and acceptance criteria. The overall score answers a broader question than most organizations are buying a model to solve.
The pattern also suggests a possible role for routing work between models. GDPx does not demonstrate that a multi-model system would outperform its individual entries, though. Routing accuracy, integration overhead, and the ability to detect a failed result would all affect whether that approach pays off.
Safety Classifiers Affect the Measured Outcome
SemiAnalysis says safety classifiers suppress some results for Opus 5.5 and GPT-6 Astra, particularly in cybersecurity, biology, and alignment. The authors also present a comparison excluding those three domains. (SemiAnalysis)
This creates an important distinction between a model being unable to solve a task and a deployed system declining to attempt it. Both can produce an unsuccessful benchmark outcome, but they imply different engineering decisions.
For a legitimate, authorized security workflow, a refusal may prevent useful work. In another setting, that same restriction may be part of the intended safety boundary. A completion score alone cannot resolve whether a restriction was appropriate.
SemiAnalysis attributes Muse Wonder Full’s strong cybersecurity showing partly to this broader issue of classifier behavior. It also suggests Gemini 4 Carbon’s biology lead may reflect weaker safety classifiers. That explanation is the authors’ interpretation, not an isolated causal result established by the leaderboard.
Excluding the affected domains changes the question being measured, too. It can help readers inspect performance outside those areas, but it removes capabilities that may matter to the buyer. I would treat the adjusted comparison as a diagnostic view, while keeping the full-domain result in sight.
Finance and Legal Remain Especially Difficult
SemiAnalysis says its finance and legal tasks are challenging enough that essentially every model scores zero. The careful reading is that these particular task sets remain largely unsolved under GDPx’s conditions. The result says little about whether a model can help with a narrower activity such as summarizing a contract or extracting figures from a document. (SemiAnalysis)
The environment can explain part of that difficulty. AI Weekly’s account of the launch describes a finance setup containing roughly 2,000 files drawn from real private-equity deal material.
A document collection of that size tests more than financial knowledge. An agent has to find the right evidence before it can reason correctly from it. A plausible answer based on the wrong document can still fail the assignment.
These near-zero domains also complicate interpretation of the overall ranking. They expose common weaknesses, but offer little separation between models that all fail. For a finance or legal buyer, the more useful next question is where the workflow breaks: retrieval, reconciliation, reasoning, execution, or final-output requirements. The headline score cannot diagnose that on its own.
Frequently Asked Questions
3 questions
1What Is GDPx?
GDPx is a SemiAnalysis benchmark containing 206 tasks across 11 specialist domains, developed with 14 data partners to evaluate frontier agents on high-value work.
