The AI industry has spent 2026 promising autonomous agents that do your job for you. A new benchmark just measured that promise — and the best model in the world completed 12% of real professional work.
That is not a typo. GDPx, a new benchmark published on November 2, 2026 by SemiAnalysis — the US research firm best known for semiconductor and AI infrastructure analysis — tested leading AI models on 206 tasks drawn from eleven real professions, from chip design and finance to law, cybersecurity, biology and software engineering. None of the thirteen models tested reached even one fifth of the total.
At TechRougeHub, we build AI automation for businesses. So we read results like this carefully — and this one deserves a critical look, not hype.
According to the published GDPx ranking, the top three models scored:
| Model | Score on real professional tasks | Average cost per task |
|---|---|---|
| Claude Opus 5.5 (Anthropic) | 12.0% | $51.41 |
| Gemini 4 Carbon (Google) | 10.9% | $16.29 |
| GPT-6 Astra (OpenAI) | 9.6% | $53.99 |
Each model ran in its native agent environment where one existed — Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon. Scores are the mean of three attempts, with every professional domain weighted equally.
No single model won everywhere. Opus 5.5 led in trading, hardware design and chips. Gemini 4 Carbon led in biology and, narrowly, software. GPT-6 Astra led in recursive self-improvement and materials. In other words: there is no "best AI agent" — there are only models that are less weak in specific domains.
Most AI headlines come from benchmarks that test answering questions. GDPx tests something businesses in the United States and Canada actually care about: can an agent finish a job — end to end, to a professional standard, without a human rescuing it?
That distinction explains the shockingly low scores. Answering a question correctly 80% of the time sounds impressive. Completing a multi-step professional task requires being right at every step — research, judgment, execution, checking — and a single weak step fails the whole task. Twelve percent is what the hype looks like when you grade finished work instead of fluent answers.
The cost column is the second story. The winning model spent an average of $51.41 per task — and over half a million output tokens per task — to achieve its 12%. Gemini 4 Carbon reached nearly the same score for roughly a third of the cost, which arguably makes it the more commercially interesting result.
For a business, the math is brutal: paying frontier-model prices for an agent that fails roughly nine times out of ten on professional work is not automation. It is an expensive experiment. The winners in 2026 are not the companies replacing staff with agents — they are the companies using AI on narrow, well-defined tasks where the output is checked and the failure cost is low.
North American companies are adopting AI agents faster than anyone — and North America also holds the largest share of the AI cybersecurity market, precisely because autonomous systems create new risk. GDPx points to a sensible adoption strategy:
GDPx does not prove AI agents are useless. It proves they are junior — fast, confident, and unreliable without supervision. The technology is improving quickly, and scores this low will not stay this low. But any vendor telling a US or Canadian business in November 2026 that agents can run a profession unsupervised is selling you the benchmark's missing 88%.
Build with AI where it is strong today — speed, drafting, pattern-finding, tireless follow-up — and keep human judgment where GDPx shows it is still irreplaceable.
TechRougeHub builds practical AI automation and chatbots with human oversight built in. Want an honest assessment of what AI can — and cannot — automate in your business?
What is the GDPx benchmark?
GDPx is a benchmark published by SemiAnalysis in November 2026 that measures how well AI models complete real professional tasks — 206 tasks across eleven professions — rather than how well they answer questions.
Which AI model scored highest on GDPx?
Anthropic's Claude Opus 5.5 scored highest at 12.0%, followed by Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%, per the published ranking.
Should my business stop using AI agents?
No. Use them for narrow, supervised, easily checked tasks — and measure how many tasks they actually finish correctly before expanding their role.
SemiAnalysis — GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains
Tech Times — GDPx Benchmark: Best Frontier AI Scores 12% on Real Work