What is the GDPx benchmark?
SemiAnalysis published GDPx on November 2, 2026. The benchmark measures how well frontier models complete multi-step agentic work in 11 domains the firm considers economically important. Version 1.0 contains 206 tasks built with 14 specialist partners.
The domains are chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Recursive self-improvement refers to an agent's ability to build the next version of itself. SemiAnalysis deliberately leaves out government, real estate and insurance.
The authors are Max Kan, Andrew Megalaa and Dylan Patel. They argue that the first trillion dollars of AI revenue will concentrate in sectors that deploy frontier intelligence at high return. They reject benchmarks weighted by GDP share or the SOC and O*NET job taxonomy. Those weights assume adoption spreads evenly.
Every task runs inside a simulated work environment rather than a single prompt. The default time limit is eight hours, though some tasks are shorter. SemiAnalysis reports pass@1 across three attempts and weights each of the 11 domains equally.
Models run in their native harness where one exists. The newsletter names Claude Code for Opus 5.5, Codex for GPT-6 Astra and Antigravity for Gemini 4 Carbon. Models without a native harness use OpenCode by default.
The task partners include Aria, Arcverex and Ergodex for chip design, Arithmetic for cybersecurity, LatchBio for biology, Living Context for finance, Chakra for legal and Third Moment for quant trading. Extrapolate supplied the RSI tasks, Beautiful Machines the hardware tasks and Datacurve the software engineering tasks.
The set is a collaboration with reinforcement-learning environment vendors. SemiAnalysis says it approached more than 30 such companies and selected one or more partners for each domain. Some domains attracted several partners that wanted exclusivity, so the firm ran a selection process based on task quality.
Which AI model tops the GDPx leaderboard?
Anthropic's Claude Opus 5.5 leads the first GDPx leaderboard at 12.0%. Google's Gemini 4 Carbon is second at 10.9%, and OpenAI's GPT-6 Astra is third at 9.6%. The gap between first and third is 2.4 percentage points.
The rest of the public ranking falls quickly. Muse Wonder Full scores 9.0%, Claude Fable 5.1 scores 8.5%, GPT 6.1 Sol scores 8.4% and Claude Sonnet 5.5 scores 8.2%. Gemini 4 Argon reaches 7.3% and Muse Wonder 6.8%.
Lower down the table, DeepSeek 4.1 Flash 1017 scores 5.4%, GPT-6 Luna 5.1%, Grok 4.7 3.9% and GLM 5.3 2.7%. SemiAnalysis says OpenAI and Anthropic generally dominate the top of the table and that open-weight models trail the leaders.
The leaderboard lists 20 models in total. Its default comparison view shows 13 of them. Muse Wonder Full appears in the ranking but has no public price, according to the site.
The spread from first to last is 9.3 percentage points. That range is wide enough to separate labs at the top, but narrow enough that a few tasks could change the order of models in the middle.
The domain wins are spread across four labs:
SemiAnalysis calls Opus 5.5 "the best all-around model" and says Gemini 4 Carbon is the only model in the same range as Opus and Astra. The authors say they want to see impressions once Google makes Carbon generally available.
Cost is a separate race. OfficeChai reports that Opus 5.5 costs $51.41 per task, roughly three times what Gemini 4 Carbon charges. The GDPx leaderboard puts Gemini at $16.29 per task. GPT-6 Astra is the most expensive of the three at $53.99 per task, so the top model delivers a 1.1-point lead at roughly three times the cost.
A 12% score means the leading model passed about one task in eight. SemiAnalysis treats that as evidence that the benchmark is far from saturated. It does not treat the result as evidence that agents are ready for unsupervised professional work.
- •Claude Opus 5.5: quant trading, hardware design and chip design.
- •Gemini 4 Carbon: biology and software engineering.
- •GPT-6 Astra: recursive self-improvement and materials science.
- •Muse Wonder Full: cybersecurity.
Why do finance and legal tasks score zero?
Finance and legal are the benchmark's weakest domains. The authors write that basically every model scores exactly zero. AI Weekly reports 0% across all 13 models it reviewed.
The finance tasks come from Living Context and are built around real transactions. One environment recreates a private-equity deal in which Bannock Capital buys about 62% of Haverly Snack Company for $880 million plus an earnout. Eakin Lowry is the sell-side banker and Marbury Kline is the accounting adviser.
That environment holds roughly 4,000 anonymized files split into four groups: a data room and three deal folders. The main data room contains about 2,000 documents. One accounting task was performed by a Marbury associate in the real transaction. The model must gather context the way a human would.
Legal tasks from Chakra follow the same environment-based approach, according to the newsletter. SemiAnalysis says it screened every partner's tasks for impossible or unfair requirements before adding them to the set.
The zero score has a commercial meaning. SemiAnalysis chose finance and legal because it expects frontier models to affect both. Each domain involves high-stakes documents and specialist judgment. A model that cannot finish one deal-room task within the default eight-hour limit is not ready to replace an analyst or an associate.
The result may also reflect the scoring design. Partial credit exists in some domains, but a finance task can fail on one wrong assumption or one missing file. The authors describe these domains as challenging enough that capable models still return nothing. That suggests a capability gap rather than a small shortfall.
How does GDPx differ from OpenAI's GDPval?
GDPx is positioned against GDPval, OpenAI's benchmark for economically valuable work. OpenAI announced GDPval on September 25, 2025, and the paper appeared on arXiv on October 5, 2025.
GDPval covers 44 occupations across the nine sectors that contribute most to US GDP. Its authors open-sourced a gold subset of 220 tasks and a public automated grading service. The paper says frontier performance improved roughly linearly over time.
SemiAnalysis argues that GDPval was designed before agents became capable. "Every GDPval task is precisely specified and non-interactive," the authors write, with all necessary context placed in the prompt. The tasks ask for a document, spreadsheet or slide deck, not a multi-step investigation.
GDPval's design reflects an earlier moment. When OpenAI published it, agents were mostly answering prompts and producing files. GDPx assumes an agent can open tools, read a file system and work through several steps before it delivers anything.
The critique also targets grading. SemiAnalysis says human experts agreed only 71% of the time on GDPval's gold subset and that the automated grader did worse. It cites vague rubrics, incorrect reference answers and links that stopped working.
GDPx replaces one-shot prompts with environments. Cyber tasks use real zero-days and running services. Materials science tasks simulate battery cyclers, a catalyst reactor and a transmission electron microscope. Hardware tasks run inside Siemens NX and test output on AWS FPGA instances.
The comparison is not neutral. SemiAnalysis sells research subscriptions and built GDPx as a rival product. OpenAI's GDPval authors have not responded inside the article. Readers should treat the critique as an argument with evidence, not a settled verdict.
Both benchmarks can be useful. GDPval offers breadth across occupations and an open gold set. GDPx offers depth in domains where agent failures are expensive. Neither replaces testing a model on your own work.
Do safety classifiers change the ranking?
Safety classifiers changed the top of the table. SemiAnalysis says Claude Opus 5.5 and GPT-6 Astra had "sandbagged overall scores due to their safety classifiers" in cyber, biology and alignment. Those classifiers can block a task before the model attempts it.
The effect is largest in cybersecurity. Muse Wonder Full, which AI Weekly describes as refusing less often, is "the best available cyber model today by a meaningful margin," according to the newsletter. Its overall score of 9.0% still trails the leaders because it is weaker elsewhere.
SemiAnalysis published a second ranking that removes cybersecurity, biology and alignment. OfficeChai reports that GPT-6 Astra then leads Opus 5.5 by 0.1 percentage points. Gemini 4 Carbon falls from second to fifth.
Biology shows the same pattern in reverse. SemiAnalysis says Gemini 4 Carbon is the best biology model, likely because its safety classifiers intervene less often. The authors present that as an observation, not a controlled finding.
For buyers, refusal rates are now a product feature. A model that is strong on paper but declines security work may be unusable for a red team. A model with looser filters may create legal or policy risk.
SemiAnalysis says safety classifiers are underrated when buyers choose a model for a specific use case. The same filter that prevents a harmful request can also block a legitimate security audit or biology analysis.
The benchmark does not resolve that trade-off. It measures task success under each lab's default behavior. Enterprises still need to decide which refusals are appropriate for their own risk tolerance.
How reliable is a 206-task benchmark?
GDPx v1.0 is small by design. Its 206 tasks divide across 11 domains, which leaves fewer than 20 tasks per domain on average. The RSI section contains 10 tasks, so one task moves that domain score by 10 percentage points.
SemiAnalysis describes a strict quality process. Partners had to pass automated QA on their first attempt, with only minor mistakes allowed. The firm also reviewed hundreds of agent trajectories by hand to check the automated findings.
The QA covered consistency, completeness, cheating and unfair verifiers. SemiAnalysis says it dropped several prominent potential partners that failed on the first pass. That screening is a strength, but it also means the published task set reflects the firm's own judgment about what counts as fair.
The authors tested that process on Terminal Bench 4.0. They write that "it flags nearly 50% of tasks as broken", which supports their claim that many public benchmarks contain unfair requirements. The number is their own QA result, not an external audit.
Several limits remain. The tasks come from partners with a commercial interest in the benchmark. The sample is small enough that a handful of tasks can reorder the table. The leaderboard also mixes models with different harnesses, prices and release status.
Availability is a specific problem. Gemini 4 Carbon is not generally available yet, according to the newsletter. Muse Wonder Full has no public price. A ranking that includes unreleased models is useful for research but difficult to use for procurement.
Replication is another open question. GDPx has not yet been independently replicated, and task leakage becomes more likely as a benchmark ages. Crypto Briefing reports that SemiAnalysis plans updates every three to six months. Updates should refresh the set but will weaken year-over-year comparisons.
None of this makes the result meaningless. A 12% ceiling across 11 difficult domains is a useful signal. The right reading is directional: frontier agents can do some high-value work, but not most of it.
What should enterprise buyers and researchers watch next?
The first thing to watch is adoption. GDPval became a reference point because major model release cards still report it, according to SemiAnalysis. GDPx needs similar uptake before buyers treat 12% as a standard measure of agentic capability.
The second is domain progress. Watch finance and legal for the first non-zero scores, and watch cybersecurity for how labs balance refusal behavior against capability. A jump in either area would change the commercial story more than a one-point move in the aggregate.
The third is cost. OfficeChai's figures show a threefold gap between the top two models per task. If Opus 5.5 keeps its lead at that price, buyers will ask whether a 1.1-point advantage justifies the spending.
Researchers should watch the task count. A version 2 with more tasks per domain would make small differences more meaningful. SemiAnalysis plans new partners and feedback cycles every three to six months, according to Crypto Briefing.
Independent review is the biggest missing piece. A benchmark built by a research firm, using tasks from paid partners, needs outside scrutiny before it becomes a procurement standard. The public leaderboard makes replication easier, but it does not guarantee it.
A public leaderboard also invites benchmark gaming. Labs can optimize for the published task styles once they are known, which is why refresh cycles and fresh partners matter.
For now, GDPx is best read as a detailed snapshot. It shows where frontier models win, where they refuse and where they fail completely. The next update will show whether those gaps are closing.