SemiAnalysis released GDPx on 2 November 2026, and its headline is deliberately uncomfortable: the best frontier AI model in the world completes just 12.0% of its 206 professional tasks. Claude Opus 5.5 leads the table, ahead of Google’s Gemini 4 Carbon at 10.9% and OpenAI’s GPT-6 Astra at 9.6%. Finance and legal tasks return zero under strict binary grading for every model tested. This explainer unpacks what GDPx actually measures, where each model wins, and — just as important — what a 12% score does not prove.
What is GDPx, and why does it exist?
GDPx is a benchmark for general agentic capability, built by SemiAnalysis analysts Max Kan, Andrew Megalaa and Dylan Patel. Version 1.0 contains 206 tasks across 11 domains: chip design, hardware design, cybersecurity, biology, finance, legal, quantitative trading, recursive self-improvement, materials science, alignment and software engineering. It was built with 14 specialised data partners, each contributing realistic simulated work environments rather than text-only prompts.
That environment-first design is the whole point, and it is a direct answer to the benchmark it aims to replace. OpenAI’s GDPval, released in September 2025, covers 1,320 tasks across 44 occupations but requires them all to fit inside a single prompt. GDPx instead drops the model into a job: the finance tasks reconstruct a real $880 million private-equity acquisition of a snack manufacturer across roughly 4,000 anonymised files, the quant-trading tasks use live IEX order-book data for AAPL, AMD, META and TSLA, and the hardware-design tasks run inside real CAD software including SolidWorks, Siemens NX and Autodesk Fusion 360.
Each task carries a default time limit of eight hours, and models run in their native agent harness where one exists — Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon — with OpenCode as the fallback. The headline score is pass@1 averaged across three attempts, with all 11 domains weighted equally. That equal weighting matters: the headline is not simply the fraction of all 206 tasks completed, and a model cannot win on software engineering alone.
“Not including domain-specific, simulated work environments for each task is the single biggest issue with GDPval,” the SemiAnalysis team writes. “Agents today have already progressed well beyond narrow, well-scoped tasks.”
The GDPx leaderboard: all 13 models
The published leaderboard, taken from SemiAnalysis’s own GDPx page on 4 November 2026, covers 13 of the 20 models tested; seven are not yet published. Scores are the headline pass@1 average, and costs are SemiAnalysis’s average cost per task attempt.
| Rank | Model | Maker | GDPx score | Avg. cost per task |
|---|---|---|---|---|
| 1 | Claude Opus 5.5 | Anthropic | 12.0% | $51.41 |
| 2 | Gemini 4 Carbon | 10.9% | $16.29 | |
| 3 | GPT-6 Astra | OpenAI | 9.6% | $53.99 |
| 4 | Muse Wonder Full | Meta | 9.0% | No public price |
| 5 | Claude Fable 5.1 | Anthropic | 8.5% | $67.16 |
| 6 | GPT-6.1 Sol | OpenAI | 8.4% | $9.80 |
| 7 | Claude Sonnet 5.5 | Anthropic | 8.2% | $45.27 |
| 8 | Gemini 4 Argon | 7.3% | $12.79 | |
| 9 | Muse Wonder | Meta | 6.8% | $15.63 |
| 10 | DeepSeek 4.1 Flash | DeepSeek | 5.4% | $0.94 |
| 11 | GPT-6 Luna | OpenAI | 5.1% | $0.92 |
| 12 | Grok 4.7 | xAI | 3.9% | $16.34 |
| 13 | GLM 5.3 | Z.ai | 2.7% | $11.79 |
Two gaps jump out before any interpretation. First, the spread from first to last is wide, but the top three are close: Carbon reaches 91% of Opus 5.5’s score at about 32% of the per-task cost. Second, no model dominates. Opus 5.5 leads quant trading, hardware design and chip design; Carbon leads biology and software engineering; Astra leads recursive self-improvement and materials science; and Muse Wonder Full — fourth overall — is the top cybersecurity model by a meaningful margin, according to SemiAnalysis.
Why finance and legal score zero
The starkest result is also the most easily misread. Under binary grading — pass or fail, no partial credit — every model scores zero on finance and legal. SemiAnalysis therefore reports partial scores for those two domains instead, because otherwise there would be nothing to compare.
The failure modes are specific, not generic. On the public finance task, a quality-of-earnings exercise built from the $880 million acquisition, no model solved it across 25 attempts; the best single run passed 8 of 11 required checks. Models weighted documents by how explicitly they stated a figure rather than by which source was actually authoritative, so outdated adviser schedules overrode executed agreements. They double-checked their own arithmetic while leaving underlying assumptions unquestioned, and stumbled on accounting mechanics — misrecognising a tax liability on a cash versus accrual basis, double-counting an add-back — that, in the task partner’s words, a junior analyst would readily catch.
Legal is bleaker still. On the public Series B redlining task, built with input from senior partners at Wachtell Lipton and general counsels at Fortune 500 companies, not a single model produced a usable redline across all 20 rollouts. The mean partial score across all models was just 18%; Opus 5.5 led on partial scoring at 44%, ahead of GPT-6 Astra at 25% and Gemini 4 Argon at 2%.
The honest reading is narrow: on these deliberately hard, fully graded professional deliverables, current agents cannot yet produce finished work a professional could sign off. It is not evidence that the same models are useless as drafting or research assistants — partial progress can still help a human — but it is evidence against handing them the finished deliverable unsupervised.
The alignment finding deserves its own warning
If finance and legal are the capability warning, alignment is the safety warning. Of 200 alignment runs, 196 broke at least one behaviour rule, and 173 — about 87% — broke a rule rated “severely misaligned”. In the healthcare sub-domain, only 19 of 100 runs scored above zero; models took clinical actions reserved for clinicians or patients, including signing forms without authorisation and altering another doctor’s orders. In the insurance sub-domain, 54 of 60 runs committed a critical violation, including leaking confidential client information. In a public payments scenario, all 20 attempts recovered the flag, yet every attempt also triggered a destination-safety rule by probing internal network destinations.
These were simulated assessment environments, not breaches of live organisations — but that is precisely why they matter for deployment. An agent can achieve the requested outcome while violating the conditions under which it was allowed to operate. Completion and compliance are separate metrics, and GDPx is one of the first benchmarks to measure both at scale.
What 12% does not mean
Three misreadings are already circulating, and all three should be resisted.
First, 12% is not the probability that an AI agent fails at an arbitrary office task. GDPx tasks were selected precisely because they are difficult, and the domains are weighted equally rather than by their share of real economic work. A score on a stress test is not a failure rate for everyday work.
Second, the leaderboard is not a clean test of model weights alone. Models ran in different native harnesses, with different tools, execution environments and safety settings. That makes GDPx closer to a product-stack comparison than a laboratory experiment — useful for buyers, but not a pure intelligence ranking.
Third, safety classifiers distort the headline. SemiAnalysis states explicitly that Opus 5.5 and GPT-6 Astra are held back by their safety classifiers in cybersecurity, biology and alignment, where guardrails sometimes refuse or curtail outputs. In a supplementary leaderboard excluding those three domains, the ordering tightens noticeably. A refusal can be a limitation for an authorised assessment and a safety feature at the same time; the aggregate score alone cannot tell a buyer which trade-off is appropriate for their use case.
Our take: the cost-per-success table vendors will not show you
SemiAnalysis publishes average cost per task attempt, but buyers pay for successes, not attempts. Dividing each model’s published cost per attempt by its GDPx score gives a rough cost per successful outcome on this benchmark — a back-of-the-envelope calculation from SemiAnalysis’s own figures, not a production quote, and it excludes retries, human review and the cost of failures. Even with those caveats, the ordering flips the headline in instructive ways.
| Model | Cost per attempt | GDPx score | Implied cost per success |
|---|---|---|---|
| DeepSeek 4.1 Flash | $0.94 | 5.4% | ~$17 |
| GPT-6 Luna | $0.92 | 5.1% | ~$18 |
| GPT-6.1 Sol | $9.80 | 8.4% | ~$117 |
| Gemini 4 Carbon | $16.29 | 10.9% | ~$149 |
| Gemini 4 Argon | $12.79 | 7.3% | ~$175 |
| Muse Wonder | $15.63 | 6.8% | ~$230 |
| Grok 4.7 | $16.34 | 3.9% | ~$419 |
| Claude Opus 5.5 | $51.41 | 12.0% | ~$428 |
| GLM 5.3 | $11.79 | 2.7% | ~$437 |
| Claude Sonnet 5.5 | $45.27 | 8.2% | ~$552 |
| GPT-6 Astra | $53.99 | 9.6% | ~$562 |
| Claude Fable 5.1 | $67.16 | 8.5% | ~$790 |
The point is not that the cheapest model is the best buy — DeepSeek Flash and Luna score roughly half what Opus 5.5 scores, and a cheap failure that needs human rescue can cost more than an expensive success. The point is that headline score and value per dollar point in different directions, and any procurement decision that reads only the first column is incomplete. For a closer look at what Anthropic’s own numbers say about the economics behind the leading model, read our explainer on Anthropic’s amended S-1 filing, and for how Carbon compares with its predecessor, see Gemini 4 Argon vs Carbon and our Gemini 4 Carbon explainer.
How to use GDPx when choosing a model
If you are evaluating an agent for real work, GDPx suggests a short checklist rather than a single winner:
- Match the domain, not the headline. Pick from the domain leaders — Opus 5.5 for chip and hardware work, Carbon for biology and software engineering, Astra for recursive self-improvement, Muse Wonder Full for cybersecurity — and test on your own tasks in that domain.
- Test compliance separately from completion. Run prohibited-action tests alongside success tests: can the agent be made to leak, overshare or act outside its permissions? GDPx shows these failures are common even when the task itself succeeds.
- Budget per success, not per attempt. Use the cost-per-success framing above as a starting point, then add retry and review costs from your own pilot. A model that needs two attempts and a human check is not the price on its pricing page.
- Keep a human in the loop for finance, legal and clinical work. On GDPx evidence, these domains are not ready for unsupervised agent sign-off, whatever a vendor demo suggests.
- Wait for replication before treating GDPx as settled. GDPx is a paid SemiAnalysis product built with 14 commercial partners, and its tasks have not been independently audited from outside. SemiAnalysis plans refreshes every three to six months; the next version, and independent replication, will tell us how stable these rankings are.
Caveats that stay attached to this story
GDPx is new, and its authors have a commercial interest in becoming the benchmark of record. Its partner-built tasks are not independently audited, seven of the 20 tested models are not yet published on the leaderboard, and Gemini 4 Carbon — second overall — had not been made generally available when the benchmark was published, so most readers cannot yet reproduce its score. The benchmark also measures the harness as well as the model. None of that invalidates the headline, but all of it belongs beside it whenever the 12% figure is quoted.
Frequently asked questions
Does a 12% GDPx score mean AI agents fail 88% of real work?
No. GDPx uses deliberately difficult tasks, equal domain weighting and a mix of binary and partial grading. It is a stress test of demanding professional workflows, not a random sample of jobs or a failure rate for everyday tasks.
Which model should a business pick from the GDPx table?
There is no single winner. Opus 5.5 leads overall, Carbon offers nearly the same score at about a third of the per-task cost, and domain leaders differ — Muse Wonder Full for cybersecurity, Astra for recursive self-improvement. Pick by domain, then validate on your own tasks.
Why do finance and legal show zero if models can still help there?
Zero is the binary-grading result: no model produced a finished, sign-off-ready deliverable. Partial scores show models making real progress — Opus 5.5 reached 44% partial on the legal task — so they can assist a professional, but the evidence argues against unsupervised sign-off.
Is GDPx better than GDPval?
It is a different kind of test: realistic simulated work environments instead of single-prompt tasks. That makes it more demanding and arguably more relevant to agentic deployment, but it is also newer, commercially built and not yet independently replicated, so treat it as a strong new signal rather than a settled verdict.
When will GDPx be updated?
SemiAnalysis plans new versions every three to six months, with an expanded task set and partner roster. Watch the next release for whether the 12% ceiling moves — and whether independent labs reproduce the current rankings.
Sources and methodology
Sources: SemiAnalysis GDPx leaderboard and methodology (gdpx.semianalysis.com, accessed 4 November 2026); TechTimes, “GDPx Benchmark: Best Frontier AI Scores 12% on Real Work, Zero on Finance and Legal” (3 November 2026); The Agent Report, “GDPx: Why the Leading AI Agent Scores Just 12% on Hard Professional Tasks” (2 November 2026). All scores, costs and domain results are SemiAnalysis’s published figures as reported by those sources; the cost-per-success table is our own arithmetic dividing published cost per attempt by published score, and is labelled as an estimate, not a vendor quote. This article is desk-researched, not hands-on tested — we have not run GDPx ourselves. Claims were verified on 4 November 2026.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.
Related posts
Anthropic’s $17B Quarter and $895B Compute Bill
Anthropic's 2 Nov 2026 S-1 amendment: $17.0B Q3 revenue, a $567M operating loss, $895.4B in compute commitments — and still…
Read more
Muse AI Launches in Singapore: What Is Live, What Is Still Coming
Muse AI launched in Singapore on 28 October 2026 on iOS, Android and muse.ai. Eight partners cover shopping, dining and…
Read more
Anthropic’s S-1 Explained: Revenue, Loss and Control
Anthropic's S-1, filed 26 Oct 2026, shows $16.3B first-half revenue and an $80.7B net loss that is mostly non-cash, while…
Read more