Theo says Grok 4.7 should not be used
A verdict with a link attached, no test, no repro, and a swipe at day-one reviewers.
Theo posted that Grok 4.7 probably should not be used at this point. That is the whole post, plus a link. No benchmark table, no codebase, no description of what he ran or what went wrong.
Grok 4.7 probably should not be used at this point
When a reply from @harjjotsinghh pushed back with a link of its own, Theo escalated rather than explained. His answer was that anyone who thought Grok 4.7 was "best model ever" at launch should not be trusted. So the argument moved off the model and onto the people who reviewed it on day one.
Anyone who thought Grok 4.7 was "best model ever" at launch should not be trusted
What is actually being claimed
Two things, and they are not the same strength. The first is a judgment about a model, offered without evidence anyone else can check. The second is a claim about the reliability of launch-day takes, which does not depend on Grok at all and is the part most engineers will find easier to agree with. Early impressions of a new model are formed in hours, usually on whatever the reviewer happened to have open, often while the provider is still tuning serving. They are weak evidence by construction.
The gap matters because the first claim is the one in the headline. Nobody in this exchange shows a failing prompt, a diff, a cost number or a task the model flubbed. If you were hoping to learn which workloads Grok 4.7 is bad at, this thread does not tell you, and nobody else in it fills the gap either.
For working engineers
Treat it as a signal about sentiment, not a result. Theo has issued fast verdicts on new releases before, including calling Gemini 4 Argon unusable after running it on his own codebases, which at least named the thing he did. This one does not even do that.
The practical version of his advice is the boring one. If a model is on your critical path, the only review worth anything is the one you run on your repository, on tasks you already know the right answer to, after the launch week serving settles. That holds whether the day-one verdict was glowing or dismissive. On the evidence published here, both are worth about the same.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.