Veo 3 launched at Google I/O just a few weeks ago, and since then we’ve seen countless videos go viral, delighting millions of people and demonstratin...
What GDPval Is Trying to Measure
Most evaluations of AI models test narrow, self-contained skills: answering trivia, solving math problems, or completing short coding puzzles. GDPval takes a different angle. Instead of asking whether a model can pass an exam, it asks whether a model can do the kinds of tasks that actual jobs are made of, spread across 44 occupations. The unit of measurement shifts from "did it get the right answer" to "could this stand in for a piece of real work."
That framing matters because the gap between a benchmark score and on-the-job usefulness is often large. A model can look strong on a leaderboard and still stumble on the messy, underspecified, multi-step deliverables that define most professional work. By sampling across many occupations, an evaluation like this tries to capture that breadth rather than rewarding a model that happens to be tuned for one narrow domain.
Why Occupational Breadth Changes the Picture
Covering 44 occupations forces the evaluation to confront how differently "work" looks from one field to another. A task in one profession might be a long written analysis; in another it is a spreadsheet, a diagram, a plan, or a piece of code. Grading that variety is harder than scoring a multiple-choice test, because the right answer is rarely unique and quality is partly a matter of judgment.
This breadth also guards against a common failure mode: optimizing a model to win on a popular benchmark while quietly getting worse at everything the benchmark ignores. When results are reported occupation by occupation, strengths and weaknesses become legible. You can see where a model already produces work a professional would accept, and where it produces something that looks plausible but would not survive review.
How to Read Results Like These
If you are evaluating a model for your own use, treat cross-occupation results as a map, not a verdict. A high aggregate number does not guarantee the model is good at your specific task, and a middling one does not mean it is useless for you. What you want to find is the slice that matches your actual work.
- Look for the occupations closest to your own, not the overall average.
- Ask how tasks were judged — whether by expert review, comparison, or automated scoring — since that shapes what the numbers mean.
- Separate tasks the model can finish end-to-end from those where it only produces a useful first draft.
- Watch for confident-but-wrong output, which is more dangerous in professional work than an obvious failure.
Turning the Signal Into a Workflow
The practical value of measuring real work is that it tells you where to put a human in the loop and where you might not need one. In areas where a model reliably produces acceptable deliverables, you can lean on it for first drafts and let a person review. In areas where it falls short, the better pattern is to use it for narrow sub-steps — research, formatting, or restructuring — while keeping the core judgment with a person.
None of this replaces testing on your own tasks. An evaluation across occupations is most useful as a starting hypothesis about where AI is ready to carry real weight, which you then confirm against the specific work you care about before trusting it with anything that ships.