Theo calls Gemini 4 Argon unusable for real code work
He says the model aggressively hallucinates about how existing systems work, and he is telling people to keep it away from their code.
Theo says Gemini 4 Argon has proven to be outright unusable for real code work in his codebases, and that the failure mode is hallucination about how existing systems work rather than weaker code generation.
In almost every test I ran, Gemini was factually wrong about over 50% of it's claims.
His number is blunt. In almost every test he ran, he says, the model was factually wrong about over 50% of its claims. He calls it a straight up liability and says to keep it away from your code. He points to a video where he goes into the methodology in more depth.
He is explicit that this is not a normal model comparison. His framing is that this is not the usual small trade where something is a few percent worse one way and a few percent better another. In a follow-up post he softens the edges slightly, saying he was genuinely hopeful for the model and that there are some good parts, but that generally speaking it should be avoided.
What is new since his first run
This is the second round of Theo testing the same model. Earlier he had it doing well on one of his benches and badly on another. What he is adding now is a verdict across his own repositories rather than a bench score, and the specific claim that the problem is factual accuracy about code that already exists.
For working engineers, the distinction matters because the thing he is describing is hard to see in a pass rate. A model that writes a broken function gets caught by tests. A model that confidently tells you the wrong thing about how your auth flow or your queue works gets caught by you, later, after you have already acted on it. A greater than 50% false claim rate on descriptions of existing systems means every answer needs independent verification, which removes most of the reason to ask.
The usual caveat applies and applies hard. This is one engineer running his own tests on his own codebases, not a published benchmark and not a reproducible harness anyone else has run. Theo has posted the methodology in video form, so the claim is at least inspectable, but nobody else has replicated it yet. Treat it as a strong signal from someone who tests models constantly, not as a settled result.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.