The research firm SemiAnalysis released a new benchmark on November 2 called GDPx, and its headline result is a useful corrective to the usual leaderboard triumphalism. The best-performing model, Anthropic's Claude Opus 5.5, scored 12.0 percent. Google's Gemini 4 Carbon followed at 10.9 percent and OpenAI's GPT-6 Astra at 9.6 percent.
Those numbers look low because the test is built to be hard in a specific way. Rather than measuring trivia or isolated coding puzzles, GDPx asks models to complete 206 tasks drawn from 11 professional domains, chosen for their economic value rather than their share of GDP. The tasks are meant to resemble the work an employee would be handed, and each model's score is averaged over three attempts per task.
Why the gap matters
There is a persistent distance between how AI models perform on abstract benchmarks and how they perform on messy, real jobs. A model can top an intelligence index and still stumble on a multi-step task with ambiguous instructions and no single right answer. We looked at the limits of ranking-as-truth when asking whether the top of the leaderboard really means the best AI, and GDPx is a pointed attempt to measure the thing the leaderboards tend to skip.
SemiAnalysis also offered an honest caveat about its own results. It noted that some models' scores were held down by their safety classifiers, which refused or hedged on tasks the systems read as risky. That makes direct comparisons tricky, because a cautious model and an incapable one can look similar on a scorecard.
Context for the frontier race
The ordering is worth sitting with. Opus 5.5 leads, but barely, and the three leading labs are separated by a couple of percentage points at a level where more than seven-eighths of the work goes undone. This lands only days after Gemini 4 Carbon put Google back on top of a more conventional intelligence index, a reminder that which model is "best" depends heavily on what you choose to measure.
SemiAnalysis says it plans to refresh GDPx every three to six months, which should make it a useful tracker of real progress rather than a single snapshot. Some observers have framed it as a successor to OpenAI's own GDPval work; SemiAnalysis describes it more plainly as an attempt at the best general test of agentic ability.
The takeaway is not that the models are useless. A system that can do an eighth of genuinely valuable professional work, reliably and cheaply, is already reshaping parts of the economy. The point is calibration. When a vendor tells you its model has mastered knowledge work, GDPx is a reasonable place to check the claim against the evidence.
Commentarii · 0