Theo times Gemini 4 Carbon at 11 seconds to first token
Head to head timings put Carbon four times slower than the competition on identical code review tasks, and Theo says almost all of it is server side.
Theo has put numbers on his complaint about Gemini 4 Carbon's speed. On the same work, he logged Gemini 4 Carbon on low at 22m21s against Opus 5.5 on high at 9m33s, and said the model is unacceptably slow in real world usage. The worst case he reported was a fix for T3 Code Windows installations, which he says took fourteen hours.
This follows his earlier early access impressions, where he called the model slow, awkward and prone to hallucination while crediting the team with real progress and some useful capabilities buried in it. The timings are the part that is new.
Where the time goes
When he looked harder at the latency, Theo put the blame on serving rather than the model. He measured 11 seconds before Carbon starts to respond, which he says is 20 times longer than Fable. He also says Varun attributes it to prefill being hard, with the dry observation that other labs appear to have figured that out.
Prefill is the pass over your prompt and context that happens before any token comes back, so it scales with how much you have stuffed into the window. In an agent loop that cost is paid again on every turn, which is how an 11 second gap turns into a wall clock number measured in hours.
Constrained to identical tasks it gets worse
Theo then narrowed the comparison so both models ran the same work, and reported Carbon at four times slower on code review tasks. That is a single engineer running his own harness, not a published benchmark, and the comparison runs against Opus 5.5 on a higher reasoning setting than the one he gave Carbon.
For working engineers, the useful distinction here is that nothing in these numbers is about output quality. Serving latency is something a lab can fix without shipping a new model, so a Carbon that is four times slower today is not necessarily four times slower next month. It also means speed complaints about a freshly launched model are weak evidence about the model itself, which is worth remembering the next time a launch week verdict arrives based mostly on how long someone waited.
Almost all of it is server side. Their inference is awful.
It's worse than I thought when limited to identical tasks
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.
