What happened, and when
On November 2, 2026, SemiAnalysis published results from GDPx, a benchmark for measuring frontier model capabilities. The headline is blunt: even the best model scores just 12%, according to the SemiAnalysis write-up.
The authors call Claude Opus 5.5 the best all-around model in the test. They also say the finance and legal tasks are so challenging that basically every model scores exactly 0. The benchmark was a collaboration with 14 specialized data partners, and the article reports that labs pay $5k+ for a single high quality task. That price tag says a lot about how expensive good evaluation data has become.
A checklist of passes and fails, the same shape as a benchmark scorecard.Why this matters for solo operators and marketers
Most of us do not run evaluations on frontier models. We buy the subscription, paste in the brief and hope for the best. A 12% headline sounds like a verdict on the models, but a benchmark measures its own test set, not your client's onboarding email or your paid social copy. The useful read is the domain split: a model that drafts well and reasons poorly about contracts is telling you where a human has to stay in the loop.
The finance and legal result is the one to pin to the wall. If nearly every model scores zero on those tasks, do not hand over a tax memo, a contract clause or a payroll calculation and assume it passed a test nobody gave it. We already draw this line on the human-in-the-loop trigger, and this benchmark is a reason to keep drawing it.
Before you pick a model off any leaderboard, build a small evaluation set from your own work. Start with ten tasks you have already done well, score the outputs, and rerun the set whenever you change models. Our piece on why your AI stack is only as good as its evaluation layer makes the case for building that layer first. If cost is part of the decision, when the cheap model is the right call explains the trade-off, and how to measure whether your AI is actually working covers what to track after launch.
Our take
Benchmarks are marketing with a lab coat on. I say that after more than 25 years in digital marketing, where nearly every platform I worked with published the metric that flattered it most. That does not make GDPx worthless. It makes it a stress test, and a zero score is more honest than most leaderboard screenshots you will see this quarter.
My position: a 12% ceiling on hard professional work is a reason to keep humans signing off on anything with legal or financial consequences, not a reason to stop using AI. The solo operators who win are the ones who know exactly which tasks a model can carry and which it cannot, and who can prove it with their own numbers. The unglamorous part is the evaluation set, and that is where I would spend your first afternoon. Nobody has ever been fired for running a boring eval, though plenty of agencies have been humbled by skipping one.
Sources
SemiAnalysis, GDPx benchmark article — https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities