I love Ultrafast (it's unusable)
Tuesday, Oct 6 · 28 min · about 8 min to read
Theo-brain watched all 28 min
Theo tested OpenAI's Ultrafast mode for the Astra model, which costs about 6x the normal price, and calls it brutally priced. Building his app Slopalytics with it, he found near-instant replies kept him in the loop and helped him design better. The lesson: speed changes how you work, but wait for a cheaper Ultrafast model before paying for it.
Skim itSkim it for the fast back-and-forth workflow ideas; the notes already cover the pricing and conclusions.
Tap any time to jump the video there.
Key takeaways
Ultrafast is brutally priced. Output tokens (the text the model writes) go from $50 to $300 per million, and $450 on long-context projects. Theo had it review two small pull requests, both under 100 lines, and the bill was $600.
is rough. a little expensive at $50 per million output tokens, but on ultrafast it goes up to $300 per million out. And God forbid this is a long context project cuz now you're at $450
mind? I see people in chat guessing like $100 per PR. It was $600 of usage for both. It's insane. It's
Speed changed the result on Slopalytics. Opus 5.5 took about 15 minutes to reach a first draft, while Astra on Ultrafast took about a minute and a half. Opus's draft looked better, but by the time it arrived, Theo had already iterated Astra's version past it.
same time. And it ended up taking the Opus run 15 minutes or so before it had a first working draft. Meanwhile, it took Ultrafast well under 2 minutes. It was close to about a minute and a half of Ultrafast work on Astra to get a
half using Astra Ultrafast. But, by the time I saw the Opus version, the Astra one was better because it only took a minute and a half and I looked at it and
The money gap is huge. His Slopalytics threads cost about $550 on Ultrafast. Normal-speed Astra would have cost $90, and Soul 6.1, which he says would have been just as good for this work, would have cost $12.
250. If I was not using ultra fast and instead I'd used the normal Astra, it would have went from $550 to 90. And if I had used Soul 61, which would have been just as good for this work, by the way, $12. So, now we have
It eats plan limits fast. A handful of small changes, including switching to dark mode and disabling two features, cost 3% of his $500 plan. That was $46.32, versus $7.72 at standard Astra pricing and $1.10 on Soul 6.1.
Now I'm down to 34 on the $500 plan. I disabled two features, I switched to dark mode, and I'm now down 3% of my $500 plan. It's insane. I'm now going to
else feels super slow. Yeah. If I was paying standard API prices for that Astra usage, it would have been $7.72, but instead it was $46.32. Yeah. Out of curiosity, I want to know what the price would have been using 61Soul instead. It would have been $1.10.
Soul 6.1 Ultrafast has been publicly confirmed as coming soon. Theo thinks it could turn the $500 plan from overpriced into a genuinely great value.
off here about ultrafast. This is a public confirmation that 6.1 Soul ultrafast is coming soon.
Either way, 61 Soul ultrafast is coming. I think it might be the differentiator that makes this $500 ultrafast plan go from feeling overpriced for what you're getting to a genuinely incredible value.
Two skills here are new to you.
In-the-loop iteration with a fast model New
- What it is
- Instead of sending one big list of changes and walking away, you keep the app and the agent side by side. You send many small requests in a row and watch each change land as it happens.
- Why Theo cares
- With Opus he looks at the output, lists 20 things to change, and waits 5 minutes to 2 hours. With a fast model, split-screening the app and the agent makes sense again, so he writes smaller requests and doesn't lose the thread.
- How to try it
- Put your running app and the agent chat side by side on screen.
- Send one small change at a time instead of a long batch, then react to what you see.
finishes. So, I found myself writing smaller requests and asking for smaller changes instead of doing what I do with Opus, which is look at the output, make a list of the 20 things I want changed, send it, and then expect it to come back in between 5 minutes and 2 hours. And
I found that it like actually makes sense to split screen with your app and your model again, because I'm not just asking it to go do the thing and then looking at it when it's done. So, if
Trim slow tool calls when the model is fast New
- What it is
- Tool calls are the actions an agent takes besides writing, such as git commands, builds or browser snapshots. When the model writes in 30 seconds instead of 10 minutes, those slow steps can double the total wait.
- Why Theo cares
- He told the agent not to touch the preview browser and only change code, because he wanted updates to show up immediately. He even wants a separate system prompt for Ultrafast.
- How to try it
- Notice which agent steps take seconds without changing your code.
- Tell the agent to skip them, e.g. 'just change code when I ask, don't drive the preview browser.'
when inference goes from 10 minutes to 30 seconds, that 30 seconds of waiting on tool calls goes from a small percentage of your runtime to doubling your runtime. So, I just told it, for example, to not make changes to the preview browser, just change code when I ask, because I don't want it to call tools that take seconds, because I just want to see the updates while we go. So,
Pick a default model by job
- What it is
- Theo switched the default model in Fleet, his repo for managing his machines, from Astra to Soul 6.1 on high reasoning with fast mode.
- Why Theo cares
- He calls it a really good balance of speed and capability for working directly on the machines in his network, as opposed to writing and deploying code.
- How to try it
- Check which model your agent is pinned to for each project.
- Use a fast, capable model for system chores, and keep a stronger model for serious coding.
machines. This is now going to be 61 Soul high fast because I think it's a really good balance of speed, capability, and performance in general when I'm trying to do tasks where I'm managing the machines in my network. Not for writing code and deploying, but for working on
Advice
Don't upgrade to the $500 plan just for Ultrafast yet. Theo says it isn't worth it, and to use Ultrafast only when you don't mind losing the usage.
things. Use it when you don't care to lose it. Don't upgrade to the $500 tier just for this yet. It's not worth it.
If you have a usage reset coming, burn the account first. Theo spends his Ultrafast allowance right before resets, because his serious work runs on Opus and an empty Codex account doesn't block him.
I'm using Opus for my serious work anyways, my Codex account being at zero is fine. It's not going to block me a whole lot. So, as soon as there is a whiff that a reset is coming, I burn that account to the ground. Sometimes the
Hot takes
OpenAI was wrong to make Ultrafast a simple toggle at this price. He has the same complaint about the max reasoning setting.
His reasoning: When you give users a knob, they turn it all the way. Prime(agen) burned over half his limit in an hour.
half his limit in an hour. They really should not have made this seem like a thing you just click. It's the same problem I have with the max reasoning level, funny enough. When you give users a knob, they will turn it all the way.
Theo can make better UIs with Astra than with Fable or Opus, even though Astra is objectively worse at design.
His reasoning: Design is about how users interact with and perceive things, so more time in the loop beats a smarter model.
It's silly to put it this way, but I can make better UIs with Astra than I can with Fable and Opus. Not because Astra is better at design, it's objectively worse, but because design is all about how users interact with and perceive
The Artificial Analysis benchmark dashboard is misleading.
His reasoning: Its headline numbers are all max reasoning, and you can't see how models in the same family relate. So Slopalytics groups models by family.
Opus 5.5 is really far on the left, even though 5.5 max is over there. It is impossible to see the relationships between models in a family in this view.
How Theo thought about this
He rejected his own cost comparison. The real choice was never paying $500 more versus waiting patiently. Without fast replies he would have lost context between steps and simply not built Slopalytics.
Because this is not a real comparison here. This is a comparison of how much Astra ultra fast costs compared to Soul 61. The difference here isn't I could have been more patient and saved $500. The difference is that I just wouldn't have built it. If I had to wait between every step, I would have lost the context of what I'm thinking about. I
He ran the numbers forward. If Soul 6.1 Ultrafast keeps the same 6x price multiplier, it would be about 10x cheaper than Astra Ultrafast and close to Astra's standard price. That's why he treats Astra Ultrafast as an experiment, not a tool.
a hypothetical world where 61 Soul comes out with an ultrafast mode that happens to be the same price multiplier specifically, not only is it 10 times cheaper than Astra ultrafast, it actually comes out around the price, if not cheaper, than Astra standard. Do you
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.
