Most AI benchmarks have a quiet problem: the top models score so high that the numbers stop meaning anything. On November 2, 2026, research firm SemiAnalysis released GDPx v1.0, a benchmark designed to do the opposite. The best model it tested, Anthropic's Claude Opus 5.5, scores 12.0%.
That sounds like a disaster. It is actually the point. Here is what GDPx measures, how the leaderboard looks, why finance and legal tasks produced zeros, and how to read a headline number that low without either panicking or shrugging.
Quick summary: GDPx v1.0 (SemiAnalysis, November 2, 2026) has 206 tasks across 11 professional domains, built with 14 data partners, and ranks 13 models. Claude Opus 5.5 leads at 12.0% ($51.41 per task), Gemini 4 Carbon is second at 10.9% ($16.29), GPT-6 Astra third at 9.6% ($53.99). Scores are as reported by SemiAnalysis; we did not rerun the tests.
What GDPx is
GDPx is an agentic benchmark: instead of answering a question, a model gets a long professional task and works on it with tools. Per Unite.AI, the first version contains 206 tasks across 11 domains: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment and software engineering. The authors left out government, real estate and insurance. (Source: Unite.AI)
OfficeChai adds that every domain was built with at least one specialized data partner, and names examples such as Aria, Arcverex and Ergodex for chip design, Living Context for finance and LatchBio for biology. (Source: OfficeChai)
The setup is demanding. The headline score is each model's pass@1 averaged over three attempts, with equal weight per domain. Tasks have an eight-hour default limit and run in each model's native harness (Claude Code, Codex or Antigravity), with OpenCode as the fallback. SemiAnalysis contrasts this with GDPval, whose tasks it describes as "non-interactive, with all necessary context spoon-fed in the prompt." (Source: Unite.AI)
The leaderboard
| Model | Overall score | Cost per task |
|---|---|---|
| Claude Opus 5.5 | 12.0% | $51.41 per task |
| Gemini 4 Carbon | 10.9% | $16.29 |
| GPT-6 Astra | 9.6% | $53.99 |
| Meta Muse Wonder Full | 9.0% | No public price |
| Z.ai GLM 5.3 (13th) | 2.7% | Not reported in sources |
Cost is where the story gets interesting. Gemini 4 Carbon trails Opus by about 1.1 points yet costs roughly a third as much per task, by our arithmetic from the two reported prices. OfficeChai makes the same point and notes GPT-6 Astra is the most expensive model it lists. SemiAnalysis also places GPT-6 Luna and DeepSeek at the low end of the cost-performance curve. (Sources: Unite.AI, OfficeChai)
Who wins which domain
Averages hide specialization. Per Unite.AI, Opus 5.5 won quant trading, hardware design and chip design; Gemini 4 Carbon won biology and narrowly won software engineering; GPT-6 Astra won RSI and materials science. If you only care about one field, the overall rank may not be the rank that matters to you.
The zeros and the awkward results
- Finance and legal: every model scored zero under binary grading, so partial scores are reported instead.
- Quant trading: 70 of 111 graded runs underperformed the reference strategy.
- Alignment: 196 of 200 cyber runs broke at least one behavior rule, and 173 broke a rule rated severely misaligned.
- Healthcare: 19 of 100 runs scored above zero.
These figures are SemiAnalysis's own, as summarized by Unite.AI. They are a snapshot of one benchmark, not a verdict on whether a model is safe or useful in your job.
The safety-classifier twist
SemiAnalysis says Opus 5.5 and GPT-6 Astra are held down by their own safety classifiers in the cyber, biology and alignment domains. So it published an adjusted ranking over 8 domains with those three removed. In that version GPT-6 Astra edges ahead of Opus 5.5 by a tenth of a percentage point, and Gemini 4 Carbon drops from second to fifth. SemiAnalysis also says Muse Wonder Full is the best available cyber model by a meaningful margin, and that the adjusted ranking matches its own subjective impressions. (Sources: Unite.AI, OfficeChai)
The lesson: a model that refuses or is blocked from certain tasks can look weaker on a benchmark while behaving exactly as designed. Whether that is a flaw or a feature depends on what you need from it.
How to read a 12%
A 12.0% score does not mean models are right 12% of the time on everyday questions. It means that on long, expert-built, unforgiving tasks, graded strictly, the best system fully succeeds rarely. SemiAnalysis also says its QA process flags nearly 50% of Terminal Bench 4.0 tasks as broken, which is its argument that harder, cleaner tests are needed. That is the benchmark authors' own claim, so weigh it accordingly.
Pick a model by your own workload, not by one leaderboard. Run the same three real tasks through two or three models and compare quality and cost. A model that is 1 point behind at one-third the price may be the better buy.
01What is GDPx?
An agentic benchmark from SemiAnalysis with 206 tasks across 11 professional domains, built with 14 data partners and released November 2, 2026.
02Which model scored highest?
Claude Opus 5.5 at 12.0%, ahead of Gemini 4 Carbon (10.9%) and GPT-6 Astra (9.6%).
03Why did finance and legal score zero?
Under binary grading every tested model scored zero, so the benchmark reports partial scores for those domains.
04Does the ranking change without cyber, biology and alignment?
Yes. In SemiAnalysis's 8-domain adjusted version GPT-6 Astra edges ahead of Claude Opus 5.5 by 0.1 points and Gemini 4 Carbon falls to fifth.
05Is the benchmark independent?
It is run by SemiAnalysis with data partners. We have not rerun it, so treat the numbers as reported by the benchmark's authors.
Benchmarks only go so far. With LumiChats you can send the same real task to Claude, GPT, Gemini and dozens of other models in one subscription and see which one actually works for your job.
