By the Smili Phone team
AI agents can write essays, answer questions, and draft code in seconds. So how good are they at real work, the kind people get paid to do? A new benchmark called GDPx, launched on November 2 by the research firm SemiAnalysis, tries to answer that question. The result is humbling. The best model scored just 12 percent overall.
That does not mean AI tools are useless. It means there is a big gap between passing a test and doing a job well, and knowing where that gap sits helps you use these tools wisely instead of trusting them blindly.
What GDPx actually tests
Most AI benchmarks look like school exams. They ask a model a question and check whether the answer is right. GDPx is different. It gives models tasks drawn from real economic work across 11 domains, including chip design, cybersecurity, biology, legal work, quantitative trading, materials science, and software engineering. The idea is to measure what an agent can actually finish, not just what it knows.
SemiAnalysis picked these domains because they carry real economic value, so success there would matter more than a quiz score.
Who scored what
By SemiAnalysis's own scoring, Anthropic's Opus 5.5 led overall with 12 percent. It won in quantitative trading, hardware design, and chip design. The firm says GPT-6 Astra led on recursive self-improvement and materials science, while Gemini 4 Carbon topped biology and software engineering. Models from OpenAI and Anthropic took most of the top ranks.
No model came close to a passing grade, and even the leaders failed most tasks. SemiAnalysis notes that safety classifiers refusing some domains lowered some scores, and leaving those domains out would reshuffle the ranking.
Why you should hold the number loosely
A single benchmark is never the full story. These are one firm's tasks and one firm's grading rubric, and there is no independent reproduction yet. SemiAnalysis also sells research to the same industry it is grading, so a healthy dose of skepticism is fair.
More importantly, a low score on frontier expert work says little about everyday tasks. Drafting a report or organizing invoices is not the same as designing a chip or building a trading strategy.
How to use AI agents well anyway
- Start with low risk jobs. Let an agent draft, summarize, or organize, then check its work before anything goes out.
- Keep a human in charge of consequential decisions. If a mistake would cost money, reputation, or safety, a person should approve it.
- Test on your own tasks. Pick ten real jobs you actually do, grade the model on those, and rerun the test whenever you change the prompt or the model.
- Break big jobs into small steps. Ask for one piece at a time, check it, then move on, instead of handing over a whole project at once.
- Watch for confident errors. A fluent answer is not always a correct one, especially on technical or legal details.
What this means for everyday users
The practical lesson is simple. Use AI agents as capable assistants, not as unsupervised experts. They are great at speeding up routine work and terrible at being left alone with work that needs judgment, verification, or accountability.
If you use AI tools for work or study in Escondido and want help setting them up safely on your phone or laptop, bring your device by the shop. We can walk through the settings with you and help you build a routine that keeps you in control.
Related: OpenAI's agent push at DevDay · Google's agent gets its own email
