Theo says Argon 4 is 10x worse, contradicting ThePrimeagen
Two engineers with their own harnesses reached opposite verdicts on the same model in a single day, and neither posted the evidence.
ThePrimeagen posted four words of praise for Google's new model, "Google cooked with Argon 4", with the usual disclaimer not to get mad at him. Theo got mad at him.
Don't get mad at me Google cooked with Argon 4
"Love you Prime but no, they did not," Theo replied, saying the model is incapable of working in a real code base without hallucinating in egregious ways. He was specific about the size of the gap rather than hedging. Not 5% worse. 10x worse.
It’s not like 5% worse, it’s 10x worse.
Neither post shows a run
That is the part worth sitting with. Both of these people build and operate their own agent harnesses, both spend their days inside real code, and both arrived at opposite conclusions about the same model within about an hour of each other. Neither one attached a log, a task, a pass rate or a diff.
So the honest read is that this is two verdicts, not two results. Theo's claim is the stronger of the two because it names a failure mode, hallucination inside an existing code base, which is at least falsifiable by anyone who wants to try it. "Cooked" is not.
For working engineers
When two competent people disagree this hard this fast, the variable is usually not the model. It is everything wrapped around it. Tool definitions, context budget, how the harness feeds files in, what it does when the model guesses. Theo has spent weeks publishing numbers on exactly those variables, including a tool-description eval where rewriting descriptions produced four times better tool use with 79% less context. A model that hallucinates egregiously in one harness and cooks in another is a plausible outcome of that gap, not a contradiction.
It also fits a pattern Theo himself has been loud about. He told people not to use Grok 4.7 and not to trust day-one enthusiasm, and he has posted a hallucination rate of over 80% for Gemini across his council review runs. He is consistent. Consistency is not the same thing as a measurement.
If you are deciding whether to point Argon 4 at your repo this week, neither tweet tells you anything you can act on. What they tell you is that the spread between "this is great" and "this is 10x worse" is currently wide enough to fit two experienced engineers inside it, and that nobody has published the runs that would close it. Until someone does, the only evidence that counts is your own codebase.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.