Argon claims a lead over OpenAI and Claude on professional tasks, but a 15% hallucination rate still needs a human check

Argon, a model we only know from a short video note, reportedly outperformed models from OpenAI and Anthropic's Claude on what the source calls the Val index. That index measures tasks of economic value. Argon scored 68.9%. The same note puts its hallucination rate at 15%, which it says is the lowest among the leading models. A hallucination is a confident answer that is factually wrong or made up.

The material we have stops there. It doesn't say who built Argon, which OpenAI or Claude versions it was compared with, how the index is built, or who ran the tests. Without that, this is a claim, not a verified ranking. It is a single score on a single benchmark, apparently self-reported. Our rule is that measurement beats marketing: a lead only counts once it holds up across hard, independent benchmarks.

Our thesis is that the second number matters more than the first. Benchmarks built around 'economically valuable tasks' are where the AI industry is heading. Labs no longer want to prove only that their model can solve puzzles. They want to show it can do billable work: reports, analysis, documents. That fits what we have seen on jobs, where back-office and administrative work is the most exposed. On that kind of work, though, accuracy decides whether a tool is usable, more than raw ability does.

That is why 15% should be read carefully. If it means roughly one wrong answer in seven, it's a real improvement over rivals, if the comparison holds, and still far too high to delegate professional work without review. A 68.9% score on valuable tasks also means the model falls short on almost a third of them. Taken together, Argon is pitched as a better assistant, not a replacement. A person still has to check its output and answer for the result.

Our reading: the race is shifting from 'smartest model' to 'most trustworthy model'. That's good news for the long term. Fewer hallucinations are what let AI move from demos into medicine, law or science, where an invented fact has a real cost. In the short term, caution applies. We will wait for independent evaluations and full methodology before moving Argon anywhere near the frontier models in our rankings. And we will keep watching hallucination rates as closely as task scores.