GDPx benchmark: what it is and how each AI model scored
2026-11-03 · 4 min read
GDPx is a new benchmark from the research firm SemiAnalysis that gives AI agents 206 long, expert-level tasks across 11 high-value fields, from chip design to quant trading, with up to eight hours per task. The best model, Claude Opus 5.5, scores 12.0%, with Google's Gemini 4 Carbon second at 10.9% and OpenAI's GPT-6 Astra third at 9.6%.
@SemiAnalysis_ launched it on November 2 with a seven-part thread, calling it "a true frontier-level benchmark: even the best model scores just 12%." The opening post has drawn more than 1.7 million views, and the labs noticed fast. Melvin Johnson, who lists himself as post-training lead for Gemini at Google DeepMind, posted that he was "super happy to see Carbon land on the Pareto frontier on GDPx," and called it a benchmark that hits the right difficulty for today's models.
What GDPx measures
The GDPx site lists the 11 domains: recursive self-improvement (RSI), chip design, cybersecurity, biology, finance, legal, quant trading, hardware design, materials science, alignment and software engineering. Each one was built with at least one specialist data partner, 14 partners in total, and SemiAnalysis handled final QA.
The domain choice is deliberate. In part three of the thread, SemiAnalysis argues the first trillion dollars of AI revenue will land in sectors that can put frontier models to work at high return, so it picked chip design, cyber and bio over government, real estate and insurance.
According to the write-up, each model ran in its own lab's agent tool where one exists: Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon, and OpenCode for the rest. The headline score averages three attempts per task, and every domain counts equally.
The GDPx leaderboard
Here are the top 10 from the public leaderboard, with the average cost per task:
| Model | Lab | Score | Cost per task |
|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 12.0% | $51.41 |
| Gemini 4 Carbon | 10.9% | $16.29 | |
| GPT-6 Astra | OpenAI | 9.6% | $53.99 |
| Muse Wonder Full | Meta | 9.0% | not listed |
| Claude Fable 5.1 | Anthropic | 8.5% | $67.16 |
| GPT 6.1 Sol | OpenAI | 8.4% | $9.80 |
| Claude Sonnet 5.5 | Anthropic | 8.2% | $45.27 |
| Gemini 4 Argon | 7.3% | $12.79 | |
| Muse Wonder | Meta | 6.8% | $15.63 |
| DeepSeek 4.1 Flash 1017 | DeepSeek | 5.4% | $0.94 |
Grok 4.7 sits at 3.9% and GLM 5.3 at 2.7%.
The overall rank hides a lot of variation by domain. SemiAnalysis says Opus 5.5 is the best all-around model and wins quant trading, hardware design and chip design. Gemini 4 Carbon wins biology and narrowly takes software engineering. GPT-6 Astra wins RSI and materials science. Meta's Muse Wonder Full is described as the best cyber model available by a meaningful margin. Finance and legal are, in the write-up's words, so hard that "basically every model scores exactly 0."
Before you quote the ranking, note that SemiAnalysis says Opus 5.5 and GPT-6 Astra lost points to their own safety classifiers, which block some cyber, biology and alignment work. Take those three domains out and the order shifts.
How to use it
You can't run GDPx yourself yet; the site doesn't describe a way to submit a model. You can still get real use out of it:
- Open the leaderboard and compare score with cost, not score alone. Gemini 4 Carbon gets within about a point of Opus 5.5 at under a third of the price per task, and GPT 6.1 Sol gets 8.4% for under $10.
- Go to the domain pages for the field closest to your work. A model that wins overall can trail in the one area you care about.
- Turn on the "Show harness" toggle to see which agent tool each score came from. A model's result partly reflects the tool it ran in.
- Look at the output token counts. Sonnet 5.5 used almost twice the tokens of Opus 5.5 for a lower score, which matters when you pay per token.
If you're comparing these models, our posts on Gemini 4 Carbon, GPT-6.1 Sol's pricing and Opus 5.5's price cut go into each one.
My read
A 12% top score on eight-hour expert tasks doesn't mean AI can't handle your invoices or intake forms. Those jobs are shorter and far easier to check. GDPx tells you two narrower things: frontier models still can't be left alone on days of specialist work, and the price of a finished task now varies by more than 50 times across the top ten.
I'd put the same two questions to any business automation project. What does a correct result cost, and who checks it? If you want to work through them for your own workflows, that's what New Face Design's free process audit is for.