Theo revises his Gemini 4 Carbon take after more tests
An early access first impression that did not survive a second round of testing, with no numbers attached yet.
Theo says he got early access to Gemini 4 Carbon and was not super impressed overall. Then, while getting ready to film a video about it, he ran a few more tests and changed his mind enough to say the video will not be going how he initially planned.
Did a few more tests as I was getting ready and...I missed some things.
That is the whole disclosure. He has not said which tests he ran, which ones he missed the first time, or what specifically the model does better than he thought. No numbers, no named tasks, no codebase.
What is actually here
This is a teaser, not a result, and worth reading as one. The useful part is the shape of the admission rather than the verdict. Theo's first pass on a new frontier model produced one conclusion and a second pass on the same model produced a different one, with nothing changing except which prompts he happened to try.
That is the normal failure mode of launch-day model takes, including the confident ones. A first-hour impression is a sample of the evaluator's habits more than a measurement of the model. Whoever reviews a model first is mostly reporting on their own test set, and the test set that gets run in the first hour is whatever was already open in a tab.
The other half of his week
Separately, Theo called Fable 5.2 the greatest model ever released and said he stayed up until 3am getting that video ready. That is a strong claim from someone who has been running Fable against Opus on real work, including the build we covered earlier, where Fable 5.2 produced a playable game in 76 minutes for $42 while Opus's version played better.
Fable 5.2 is the greatest model ever released.
So the running scoreboard right now is one superlative with hours of hands-on work behind it, and one revised impression with nothing public behind it yet. Wait for the Gemini 4 Carbon video before updating on Carbon, and treat anybody else's day-one verdict on it with the same discount.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.