It’s so much worse than I thought…
Friday, Nov 13 · 20 min · about 7 min to read
Theo-brain watched all 20 min
Theo follows up his Gemini 4 Carbon review to say he was too soft on it. When he audited his own agent runs, it was about five times slower than other models, largely because every step waits over 11 seconds before it starts working. His conclusion: don't use Gemini for coding work right now. A $20 Claude subscription gets you more done.
Briefing is enoughThe notes cover the numbers and reasoning that matter, so you can skip watching unless you want to hear Theo make the case himself.
Tap any time to jump the video there.
Key takeaways
Theo audited his real usage logs and found Gemini 4 Carbon was consistently about five times slower than any other model on comparable work. He calls this much worse than he said in his first review.
done. So I took the time to audit my real world usage of all of these models. And in like for like work, Gemini was consistently five times slower than anything else. However bad you think this is,
Most of the delay comes before the model does anything. Each request waits over 11 seconds before it starts working. Coding agents work in many steps, and Gemini takes more steps than other models, so that wait stacks up until the model becomes unusable.
the request even lands. This model consistently takes over 11 seconds for when the request hits their server to when it starts doing something. 11
step. When you combine this with the fact that Gemini takes more steps to complete tasks and you have to compound this additional 11 seconds per 20 seconds on every single request, the result isn' just embarrassing. It's
Theo calls the Artificial Analysis speed chart misleading, even though it already ranks Gemini as the slowest. It only times from when Google's server accepts a request, and many of its tasks are single-step. Real coding agents run many steps, and that is where Gemini crumbles.
has completed. This is wrong and misleading because this measures from the time the request officially is hit and accepted by the server on Google side to when the response comes in. And
a lot of the tasks for artificial analysis are not agentic, so they're not multi-turn. Once you get into multi-step turns, which is, guess what, how every single agentic coding tool works, this crumbles so fast. As I mentioned
Google's caching is broken. A cache stores work the model has already processed so it doesn't redo it. Theo says Gemini has about a 30% chance of not reading the cache it just wrote. It may read a stale one from 30 minutes ago or none at all.
are the only lab where you write to cache and then read from cache and it's a 30% chance it doesn't read from the cache you just wrote. It'll read from something. It might be a cache from 30 minutes ago. It might just fail to read
The only real upside Theo sees is that Gemini is slightly less likely to flag you when you do cybersecurity work. Meanwhile the price doubles on December 1st.
been. I cannot respect the engineering opinions of anyone who thinks this model is usable for real-world engineering work outside of the fact that when you do cyber security stuff with it, it's slightly less likely to flag you. There
And of course, the subscription subsidy side, it is so disgustingly expensive too, especially starting December 1st when they double the price. Yeah, this model's going to double in price in 3
One skill here is new to you.
Tell agents to stop waiting on a slow helper model New
- What it is
- Theo's agents brought Gemini in as a 'judge' to review their work. He updated his agent skill (a reusable instruction file) to say: Carbon is slow, so if you're stuck waiting for it and have work to fix, stop it and get back to work.
- Why Theo cares
- Carbon was so slow he didn't want his agents waiting on it. He also notes the change gave Gemini an advantage in his numbers. Without it, its measured speed would have looked even worse.
- How to try it
- Pick a slow model or step your agents routinely wait on.
- Ask your agent to add a line to its skill: if stuck waiting on that model while other work remains, stop it and continue.
- When you later judge that model's speed, remember the cut-short runs make it look faster than it is.
it seems. I had nine separate times that Carbon was brought in as a judge for work. And you'll notice after the first three,
It was so slow that I went and updated my skill to tell my agents, by the way, it looks like carbon's pretty slow. If you're stuck waiting for it, but you have work to fix, don't bother, just stop it and get back to work. If I hadn't done that, the numbers would have been worse. I accidentally gave them an
Advice
Skip Gemini and get a $20 Claude subscription instead. Theo says you'll get more inference, more value and more work done.
function properly. But until then, seriously, just just go get a $20 Claude sub. You're going to get more inference out of that, more value out of that, more work done. There is no reason to
Don't take Google seriously as a company for running AI models (inference) until these problems are fixed.
fundamental level that Google does not give a single flying about anyone using their inference. And until these things are fixed, you should not take them seriously as a company. And I hate
Hot takes
Theo says he can't respect anyone who picks Gemini right now. He has lost respect for some engineers he thought knew better who publicly claim it's good for day-to-day work.
His reasoning: He says it is simply that bad.
alternative to other things. I can't respect you if you pick Gemini right now because it is that bad. And I am sorry
is literally no reason otherwise to use this model right now, and I have lost respect for a handful of engineers that I thought knew better, that are publicly claiming it's good for day-to-day work.
Gemini isn't 5–10% worse than the leaders. Theo says it is an order of magnitude worse in almost every scenario.
His reasoning: A reply Theo calls a 'Varun-ism' said there's always a tradeoff between latency and cost. Theo answers that in the screenshot being discussed, Gemini is the most expensive, the worst and the slowest model all at once.
thing. One last veroonism before we wrap here. We're We're into how we can optimize this more. There's always a tradeoff of latency and cost, but we should do better here. This is the type
realize that the model in the screenshot is the most expensive in the screenshot, the worst in the screenshot, the slowest in the screenshot, and the most miserable to work with. This is effectively admitting that they cannot keep up on any access. This no longer is
You'll know Google has finally made a great model when its own products and infrastructure stop falling apart.
His reasoning: A model only counts as good enough once it's smart enough that Google can point it at its own internal systems and fix all of its broken stuff.
it would be usable is if the model gets smart enough that they can set it on their internal and fix all of their broken stuff. But until they do that, I've said this before and I'll say it again, you'll know that Google has finally made a great model when their products and infrastructure stop falling
How Theo thought about this
Theo ruled out the obvious explanations first. Slow output speed, too little compute and badly run parallel tool calls are all real, but together they explain only 30–50% slower, not 5x. That led him to look earlier, at the wait before the request even starts.
properly. All those things are true, and all of those would make it like 30 to 50% slower. You don't get 5x slower by doing tool calls wrong. Okay, you can get 5x slower by doing tool calls wrong. They're not doing them that wrong. They're close. The problem comes before the request even lands. This model
Theo was confused about why Gemini was still so slow. He pointed to his video on OpenAI moving Codex to WebSockets, a live connection that avoids resending the whole history every step. The reply said that probably wouldn't immediately solve it. Theo conceded the point, kept digging and found much crazier numbers.
was and it was. So I was confused how it was still so slow. So I recommended a video I had done in the past, which is about OpenAI moving to WebSockets for inference in Codex. The reason they did
this. And to his credit, he's right. This probably wouldn't immediately solve it, but I was annoyed enough about this that I chose to dig in more. And that's when I found these much, much crazier numbers. Gemini 4 Carbon takes more time
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
