The best artificial intelligence model in the world can complete about 12% of a set of long, realistic professional tasks. That is the headline of GDPx, a benchmark the research firm SemiAnalysis released on November 2, and it is less a verdict on AI than a reality check on how AI is measured — and what happens when the test stops grading answers and starts grading work.
A test that grades jobs, not exams
Previous broad benchmarks mostly handed a model a prompt and a document and asked for a deliverable, graded against a rubric. GDPx drops the agent inside a simulated workplace instead: a reconstructed private-equity deal, live cybersecurity services, chip-design tools and lab instruments, built with 14 specialized data partners. Tasks carry a default time limit of eight hours, and the headline score averages each model’s first-attempt pass rate across three attempts, with every domain weighted equally.
Version 1.0 contains 206 tasks across 11 professional domains: chip design, cybersecurity, biology, finance, legal, quantitative trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. SemiAnalysis deliberately left out industries such as government, real estate and insurance, focusing on sectors it expects to generate high returns from frontier AI.
The leaderboard
| Rank | Model | Developer | GDPx score | 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% | Not publicly priced |
| 13 | GLM 5.3 | Z.ai | 2.7% | — |
The top of the table is tight: only 1.1 percentage points separate Opus 5.5 from Gemini 4 Carbon, and Carbon costs roughly a third as much per task. Opus 5.5 led overall and won quantitative trading, hardware design and chip design, according to SemiAnalysis’s own summary; GPT-6 Astra led on recursive self-improvement and materials science, and Gemini 4 Carbon topped biology and software engineering. SemiAnalysis notes that Gemini 4 Carbon is not yet generally available.
Why the numbers are so low
Jobs have a property exams don’t: you can get 90% of the way there and still have accomplished nothing, because the deliverable is wrong, a rule was broken or a file was never found. That is why a 12% score is not a story about models getting worse — it is a story about the test getting honest.
Two caveats matter. First, finance and legal tasks were so hard that nearly every model scored zero under strict pass/fail grading, so SemiAnalysis published partial scores for those domains instead. Second, the firm says safety classifiers pulled down some models’ scores in cybersecurity, biology and alignment — and that excluding those three domains moves GPT-6 Astra from third to first. Models also ran in their native product harnesses where available — Claude Code for Opus, Codex for Astra, Antigravity for Carbon — so GDPx compares model-plus-harness systems, not bare models.
What it does not prove
A 12% score on eight-hour frontier tasks does not mean a model fails 88% of short, well-defined ones. It says nothing direct about tasks outside its 11 domains, and nothing here has been independently replicated: GDPx is a SemiAnalysis product, its partner-built tasks are largely private, and every number in the launch leaderboard is SemiAnalysis’s own reporting. The firm plans new versions every three to six months, so these numbers will move — and other labs will now have a strong incentive to beat them.
That distinction is the useful one for anyone deciding what to hand an agent at work. Short, well-specified tasks with clear checks are a different job from an eight-hour workflow in a messy environment — and GDPx is, for the first time at this scale, measuring the second kind.
What to watch next
Watch whether the next GDPx release narrows the gap between the leaders, whether rival labs publish independent replications, and whether any model breaks through on finance and legal tasks, where strict grading currently defeats nearly everyone.
Sources: SemiAnalysis GDPx announcement (Nov 2, 2026), via Unite.AI, “SemiAnalysis Launches GDPx Benchmark Spanning 11 Agentic Domains” (Nov 2, 2026); Dev.to, “12% of the Economy: The Benchmark That Grades AI Agents Like Employees” (Nov 2, 2026); Construction AI Brief summary of Unite.AI reporting (Nov 2, 2026).
