Jev Daily

Theo clocks Gemini at 11 seconds before first token

One developer's timing run, no harness published, and the slowdown holds when the tasks are matched.

Theo dug into Gemini's speed and says the delay is not on his end. By his measurement it takes 11 seconds before the model starts to respond, which he puts at 20x longer than Fable. His read on where the time goes is blunt. "Almost all of it is server side. Their inference is awful."

Theo - t3.gg
@theo
X
It takes 11 seconds before it starts to respond, 20x longer than Fable.
Nov 11, 2026 · View on X

He then narrowed the comparison to identical tasks, on the theory that a chunk of the gap might be the two systems doing different amounts of work. It got worse rather than better. On code review tasks with the work held constant, he reports Gemini running 4x slower.

Theo - t3.gg
@theo
X
literally 4x slower in code review tasks
Nov 11, 2026 · View on X

What this is and is not

This is one developer timing two systems on his own workload. Theo did not publish the harness, the task list, the number of runs, or the region he was hitting. There is no variance figure and no indication of whether the 11 second figure is a median or a single observation. Time to first token is the metric that matters most for anything interactive, and it is also the metric most sensitive to where you are sitting and what time of day you run it. Treat the numbers as a direction, not a benchmark.

The second post is the more interesting one. Raw wall clock comparisons between coding agents usually tell you as much about how much work each one decided to do as about how fast the inference is. Theo controlling for the task and still landing on 4x is the part that would be hard to explain away as one agent simply doing more.

For people building on it, note what was measured. This is coding and code review work, not the decisions endpoint, so do not carry an 11 second start or a 4x multiplier over to latency expectations for Gemini Decision 1. Those are different serving paths and nothing here measured them. If start latency is load bearing in your product, the useful move is to run your own timing against your own prompts in your own region before you take a number off a timeline, including this one.

Know what Jev is good at before you bet on it

One email a day. Under three minutes. What builders found out about Jev yesterday.

Get the next one by email

Jev, read daily so you do not have to. The builds, the benchmarks, the criteria that worked and the cases where it lost, from the people shipping on TypeSafe AI's System One model.