It's Here.
Friday, Sep 4 · 44 min · about 9 min to read
Theo-brain watched all 44 min
Theo reviews OpenAI's GPT-6 Astra and calls it the best model ever. It is far ahead at computer use (the model driving your mouse and keyboard) and 3D work, keeps code changes in scope, and costs under half of Anthropic's Fable per task because it uses fewer tokens. Anthropic still wins at design, and Astra sometimes agrees to fix review comments and then doesn't.
Skim itWatch the 3D game demo, the front-end comparison, the Lakebed results and the frustrating PR thread; the notes cover the benchmark run-through.
Tap any time to jump the video there.
Key takeaways
On computer use (the model driving your mouse and keyboard), Theo says nothing comes close; it feels like two or three generational leaps. On the OSWorld test Astra scored 71.6% in about 23 minutes. Soul's best was 65.7% and took almost an hour and 15 minutes.
just need you to trust me. If you care a lot about computer use, there is nothing even close. This model is not just like a single generational leap in computer use. It feels like two or three. It's absurd. And it makes anthropic models
computer use tasks. On high, Astro was able to complete OSWorld 2.0 in about 23 minutes with a 71.6% score. Soul's best score for reference was a 65.7% and it took almost an hour and 15 minutes to complete. That's a
Price is $10 per million input tokens and $50 per million output, almost the same as Anthropic's Fable. Astra uses far fewer tokens, so the cost-per-task numbers Theo showed put it at under half of Fable's cost, and cheaper than Opus.
First off with the cost. The price is nearly identical to using Fable in terms of the tokens in and out at $10 per million in and $50 per million out.
metric. And you can see here that Astra comes out to under half the cost of using Fable for the same tasks and cheaper than Opus for the same tasks as well. Those are some good numbers.
For code it beats Soul by enough that Theo can now fairly confidently merge its changes. Soul wrote tests nobody needed and went too far. Astra is more restrained and keeps changes in scope.
is unbelievably better than what I was getting out of 5.6 Soul to the point where I can actually somewhat confidently merge changes from this model where that was not the case previously when I was using Soul. I
made sense, writing tests that didn't need to be there at all and just going too far in places it probably shouldn't. Astra is much more restrained in those ways and seems to understand the scope of the changes is making better overall.
Theo still prefers Anthropic for design. Astra struggled to mock up interfaces he could actually use. Fable got mockups close to what he wanted within one or two prompts.
design stuff and I had a ton of problems trying to get this model to mock UIs that I could possibly actually use where with Fable I was able to have it come in and get mocks pretty much exactly where I wanted within one or two prompts. I
Rough edge: in one PR thread Astra agreed review comments needed fixing, then didn't fix them. This happened even after he had corrected it earlier in the thread and had his instructions loaded. A newer version is better, but it reduced this behavior rather than removing it.
fact that not only had I corrected this behavior earlier in the same thread, I also had the skill in context. It knew exactly how I wanted these things to be handled. It instead of doing that said, "Yeah, I should do that." And then didn't. At which point I said, "Well,
better specifically at these edges. But these things do still happen. It was a reduction in bad behavior, not a removal of it. So, I wanted to showcase this
Three skills here are new to you.
A "babysit" skill for pull requests New
- What it is
- A written instruction file (a "skill") telling the agent to watch a PR (a proposed code change). It fixes failing automated checks, rebases (updates the change onto the latest code) when it conflicts, and answers review-bot comments. It keeps going until everything is green.
- Why Theo cares
- It spells out exactly what he wants done with a PR. Even so, in his test Astra stopped monitoring too early, and he later found more comments it had missed.
- How to try it
- Ask your agent to write a babysit skill: monitor the PR, fix CI failures, rebase on conflicts, address review comments.
- State that it must not stop until every check is green.
- Invoke it whenever a PR is filed, then spot-check whether new comments slipped through.
skill that I wrote that explicitly explains to the model what I want it to do. I want it to keep an eye on the PR, usually through polling or through some monitoring tech. I want it to address CI
conflicts, rebase it. And most importantly, I want it to address comments as they come in, in particular, from those review bots. And it should not stop monitoring until everything is a check mark in green. And this is where
Two-prompt audit, then fan out to sub-agents
- What it is
- First ask the model what it would improve before launch. Then tell it to spin up sub-agents (parallel helper agents) to do it all, with permission to merge when satisfied.
- Why Theo cares
- On his Lakebed cloud project this cut sync delay from as high as 800 milliseconds to under 30. In many cases it cut slow-case latency by 98%.
- How to try it
- Ask your agent: what should we improve or focus on before launch?
- Reply: spin up sub-agents and go do it.
- Only allow auto-merge if you will test the results yourself afterward.
first prompt was me asking if it thinks there's anything we should improve or focus on in Lakebed before launch. And then the second one was, okay, cool. Spit out some sub agents and go do it. And it did. And I told that it could
changes as they occurred. Astra got that time down to under 30 milliseconds. In many cases, it shaved the P95 by 98%.
3D scenes and games via Blender New
- What it is
- Give the model access to Blender, the 3D modeling tool, and ask it to build a scene or game.
- Why Theo cares
- He says it is so far ahead at 3D that comparing it to other models is unfair. But his fish game's controls came out janky: mouse too fast, then too slow. That took back-and-forth to fix.
- How to try it
- Have your agent connect the model to Blender and describe the scene you want.
- Expect great visuals on the first try.
- Fix the feel by telling it what you dislike and how to change it.
this model does. In particular, when you give it Blender access and tell it to create something like this, it's so far ahead in this type of 3D design work that it's unfair to even compare to
not get the controls right at all. Movement sucked and was janky. The mouse movement in particular was way too fast initially, so I told it to slow it down a bit. Then it made it way too slow. All
Plug it into Gmail and Notion to triage your week New
- What it is
- Theo connected the model to his Gmail and Notion and asked it to find what he should be prioritizing.
- Why Theo cares
- With his assistant out, it surfaced so much he had missed that it made him feel bad about how far behind he was.
- How to try it
- Connect your agent to your email and notes tool.
- Ask it to list what you should be prioritizing and what you have dropped.
- Work through the list it returns.
life right before filming because I plugged it into my Gmail and notion and I had it help me find things I should be prioritizing. Admittedly, my assistant is out this week, so it's all been on me. So, I fell behind on a lot. It did
Advice
Use AI code-review bots on your PRs. Theo says he couldn't live without them, and on this PR they caught real problems.
know if any of them are sponsoring this video, but I've said many a time I couldn't live without the AI review spots. And this is another great example of why it had real findings. So, I just
Decide where your trust bar is: the point where you stop checking the model's work constantly. Then keep testing models against it. Theo says he is past that point for many tasks.
where is your bar? At what point are you going to stop checking the work the model does constantly and let it do its thing? I know I am past that point in so many of the things I do, but I'm curious
Hot takes
GPT-6 Astra is the best model ever.
His reasoning: It has so many unique capabilities that nothing else comes close.
video. Is it the best model ever? I think I'm comfortable saying yes there. This model has so many unique capabilities that nothing else comes close to that it's an easy cell for me to say that. It's just insane. It makes
Artificial Analysis, a popular benchmark ranking, is not a great benchmark anymore.
His reasoning: It mixes many tests of different ages, and the older ones don't show what models do now.
it. This is not a great bench. My suspicion as to why is because it's combining a bunch of different benches of varying ages. And a lot of the old benches just aren't really showcasing what models are doing now. None of these
Astra can't actually edit video, despite the hype.
His reasoning: The YouTube shorts it cut for him were bad. The people praising its editing aren't credible editors, judging by their own channels.
all cringe and bad. I don't think this model can actually video edit. And when I went and looked at the people who were saying it were, and then I checked their YouTube channels, no offense, they are not worth trusting when it comes to
Legacy apps that have rotted for years can now be replaced in days.
His reasoning: His Plex replacement took about five prompts to become his main player and about eight to get well ahead of the competition.
doing other work. It took like five prompts to get this to the point where I would want to use it as my main player and like eight to get it genuinely far ahead of the competition. It's silly. All these legacy apps that have been rotting for years can now be replaced in
How Theo thought about this
He let Astra merge its own Lakebed changes but didn't trust it blindly. Soul had burned him before with unchecked merges, so he tested every change himself, and everything was good. He credits that to the model testing its own changes and coordinating swarms of agents to verify its work.
little nervous of these changes because I've been bit so hard by letting soul yolo merge in the past. So, I thoroughly tested all the changes it made here, and everything was good. A lot of that comes from the model's ability to test its own changes and to coordinate swarms in order to verify the work it's doing more
He separated OpenAI's demos from his own results. The slideshow demos looked great, but his own slides were worse. He allowed that his prompting might be partly to blame, yet still judged the demo unrepresentative.
incredible. But for my experience, asking it to do similar things. The computer use side was crazy. Like watching it actually navigate my computer was nuts. But the quality of the slides it generated just was not as good as what I'm seeing in these. I'm
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.
