Modern Creator
Matt Wolfe · YouTube

OpenAI GPT-6.2 Astra Is Here, and the Real Story Is Self-Improvement

OpenAI shipped two new models that are cheaper and more efficient, not smarter, and buried in the announcement is the first public model to cross its high threshold for AI self-improvement.

Posted
3 days ago
Duration
Format
Review
educational
Views
121.9K
1.7K likes
Part of the collectionThe GPT-6 Astra PlaybookEvery GPT-6 Astra breakdown, synthesized into one page.
Read the playbook
Big Idea

The argument in one line.

GPT-6.2 Astra is a cheaper, more token-efficient model rather than a smarter one, and the bigger story is OpenAI saying it now meaningfully speeds up the people building the next AI.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • You pay for ChatGPT or the OpenAI API and want to know whether switching to Astra or Sol is worth it over Claude Fable or Opus.
  • You follow AI news and want a plain-English read on what OpenAI means by the high threshold for AI self-improvement.
  • You build with Codex or Claude Code and want to know whether coding scores moved with this release.
SKIP IF…
  • You already read the OpenAI system card and the early-tester threads and just want raw benchmark tables.
  • You want a hands-on tutorial for using Astra; this is a news and reaction breakdown with a handful of quick tests.
TL;DR

The full version, fast.

OpenAI released GPT-6.2 Astra and Sol. Astra costs $8 input and $40 output per million tokens, 20% below the last Astra and equal to Claude Fable 5.2, but it uses far fewer tokens per task, about $4 versus $26 on Artificial Analysis. It ranks second at 62 behind Fable's 65, and coding scores did not move. Early testers praise 3D, design and computer use but say it stops early and is hard to follow. The headline claim is that Astra is OpenAI's first public model at the high self-improvement threshold, meaning it speeds up AI researchers, though progress is still on the same trend line as before.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 00:58

01 · Two models and a buried line

Astra and Sol launch, with OpenAI claiming the first public model at the high self-improvement threshold.

00:58 – 01:56

02 · Pricing and availability

Astra at $8/$40 per million tokens, 20% cheaper and equal to Fable 5.2; Sol unchanged at $2/$10; live in ChatGPT, Codex and the API.

01:56 – 03:09

03 · Benchmarks: second place, far cheaper

Artificial Analysis ranks Astra second at 62 behind Fable 5.2 at 65; about $4 versus $26 per task; coding flat.

03:09 – 04:37

04 · What early testers said

Theo, Matt Shumer, Dan Shipper's team, Ethan Mollick and Matt Berman: strong at 3D and review, but stops early and is hard to follow.

04:37 – 06:11

05 · Where Astra shines and Busey Bench

OpenAI's 3D and design demos plus Gary Busey SVG results from Astra, Astra Pro, Sol and Sol Pro with token and cost figures.

06:11 – 08:00

06 · Megabonk clone test

Wolfe clones Megabonk with Astra: great opening screen and working controls, skateboard facing the wrong way.

08:00 – 09:40

07 · What high self-improvement means

The system card evals: research debugging, kernel writing, NanoGPT speedrun; Astra cuts training time 43% versus 72% for the best human.

09:40 – 10:33

08 · Risk, Control Room and Anthropic's opposite call

Why faster AI-built AI raises safety risk, OpenAI's Control Room, and Anthropic holding back Fable 5.2 in the same area.

10:33 – 13:00

09 · Bottom line

A decent, cheaper model that is not a game changer; stay on Claude if you are happy, and watch the self-improvement story.

13:00 – 14:58

10 · Editor Matt: a vibe-edited video

Wolfe reveals Fable 5.2 edited the whole video and notes the recognizable look of AI-edited videos.

Atomic Insights

Lines worth screenshotting.

  • OpenAI's pitch for GPT-6.2 Astra is better results with fewer tokens, so the same job costs less than it did on the previous model.
  • Astra's API price of $8 input and $40 output per million tokens matches Claude Fable 5.2, yet a benchmark task costs about $4 on Astra against about $26 on Fable.
  • On Artificial Analysis, Fable 5.2 still leads at 65 and Astra is second at 62, a five point jump over the last Astra.
  • Coding did not improve: Astra in Codex scores 53, the same as before, and Sol dropped three points.
  • Nobody called Astra dumb. Early testers described it as hard to work with: too short, stops early, and hard to follow.
  • 3D, design and science are where Astra shines, and even its critics agree on that.
  • OpenAI defines the high self-improvement threshold as the model helping researchers about as much as an extra capable research engineer for each of them.
  • On training a small language model faster, Astra cut training time by about 43 percent, while the best human result is about 72 percent.
  • OpenAI says its models are not yet improving faster than before; this release sits on the same trend line as GPT-5.
  • OpenAI built a central stop button for its research systems called Control Room because AI that builds AI can ship new risks faster than safety work keeps up.
  • Within a week, Anthropic held back Fable 5.2's AI-development skills while OpenAI shipped Astra with more monitoring, opposite calls on the same question.
  • The question to ask of new models is shifting from whether they are smarter to whether the people checking them can keep up.
Takeaway

Cheaper per task, not smarter, and self-improvement is the story

WHAT TO LEARN

Astra's win is efficiency, not intelligence, and the lasting signal is that frontier labs now say their models meaningfully speed up their own researchers.

  • Same per-token price as Claude Fable 5.2 can still mean a far cheaper job when a model uses fewer tokens, about $4 versus $26 per benchmark task here.
  • Judge a release by the benchmark that matches your work: Astra moved up overall and in science and 3D, but coding scores stayed flat.
  • Early-tester complaints clustered around behavior, not capability: stopping early, asking unanswerable questions, and compressed explanations.
  • Day-one reactions come from pre-release builds, so test a new model on your own small set of tasks before changing your default.
  • High self-improvement means the model speeds up AI researchers, such as cutting a small-model training run by 43 percent, not that ChatGPT is rewriting itself.
  • Progress is still on the same trend line since GPT-5, so this is a milestone toward faster loops rather than a runaway one.
  • Labs are making opposite calls: Anthropic held back AI-development skills in Fable 5.2, while OpenAI shipped Astra with added monitoring and a Control Room.
  • If you are happy on Claude Fable or Opus, Wolfe sees no reason to switch; try Astra on visual work, and drop to Sol if it feels too terse.
Glossary

Terms worth knowing.

Astra
OpenAI's top GPT-6.2 model, built for the hardest projects and positioned as the rough equivalent of Anthropic's Claude Fable.
Sol
OpenAI's cheaper everyday GPT-6.2 model, positioned as the rough equivalent of Anthropic's Opus. It costs $2 input and $10 output per million tokens.
High threshold for AI self-improvement
OpenAI's label for a model that helps its researchers about as much as giving each one an extra capable research engineer. It does not mean the model rewrites its own code.
Artificial Analysis
A site that aggregates many benchmark scores into one index to rank models by overall intelligence.
Control Room
OpenAI's central stop button for its research systems, built as a safeguard now that models help build the next models.
Busey Bench
Matt Wolfe's own test where models draw Gary Busey as an SVG, scored by an LLM judge.
Resources

Things they pointed at.

01:57toolArtificial Analysis
05:20toolBusey Bench
06:40productMegabonk
Quotables

Lines you could clip.

12:38
“The models are starting to help build the models.”
one-line thesis for the whole video→ TikTok hook↗ Tweet quote
08:50
“It's AI that meaningfully speeds up the people building the next AI.”
plain-English definition of self-improvement→ newsletter pull-quote↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

OpenAI just released two new models, GPT 6 .2 Astra and GPT 6 .2 Sol. And sort of buried down in the announcement was this line. They're saying this is their first public model to hit what they call the high threshold for AI self -improvement.
At the same time, the people who tested it early are not like super excited about it. So in this video, we're going to cover what's actually new, what people are saying, a couple of my own tests, and what the self -improvement line really means for you.
So first, let me go over the basics. Astra is OpenAI's top model, the one for your hardest projects. This is like their equivalent to Fable from Anthropic.
And then Sol is the slightly cheaper everyday model. If you're looking at Anthropic's model, it's kind of their Opus equivalent. If you're using it through the API, Astra is $8 per million tokens input and $40 per million tokens output.
That's actually 20 % cheaper than... the last version of Astra. And it's also exactly what Anthropic charges for their Claude Fable 5 .2.
Now, Sol is staying the same at $2 input, $10 output. And both are live today inside of ChatGPT work and Codex if you're on one of the paid plans. So Plus, Pro, Enterprise, or Edu.
And they're also available in the API right now as well. Now, this time around, OpenAI's pitch isn't really that this model is like way smarter. They're more claiming that you're gonna get better results.
using fewer tokens. So the same jobs are gonna cost less with this new model than they would have with the previous model. Now, super quick disclosures, I did not get early access to this model.
So I'm gonna be leaning on some of the community members that did and talking about some of their opinions, plus some of my own tests. And also yes, OpenAI has sponsored this channel in the past, but this video is not sponsored by them. All right, so first let's check artificial analysis.
like an aggregated benchmark that takes a whole bunch of benchmark scores, matches them together and tries to find the objectively smartest model. And on artificial analysis, Claude Fable 5 .2 is still number one at a score of 65. The new Astra model now comes in second at 62.
So a five point jump from the last Astra model. It's also now ahead of Gemini 4 Carbon and Claude Opus 5 .5. And it also comes in first on their science and 3D test.
But again, the big narrative here is more about efficiency than it is smarts. Running this benchmark costs about $4 per task on Astra and on Fable, it's about $26. And if you remember when we looked at the API costs, they're the same cost.
This new model just uses way less tokens. It's just a much more efficient model. However, when we look at coding, nothing really moved.
Astra inside of Codex scores 53. So, you know, same as the last. Sol actually dropped three points.
Fable inside of Claude Code is at 58. So on paper, it's the second best overall. It's a lot cheaper to run, but it's no better at coding, which is actually pretty solid when you consider all of those details.
But it also makes the reaction that everybody's having to this model a little bit surprising. Theo from T3 posted this letter that he sent OpenAI during his testing. He says that it's great at reviewing code, great at 3D and games, and computer use is nuts.
But he also says he wouldn't pick it over Opus 5 .5. It stops in the middle of long tasks to ask questions he can't answer, and he can't follow its explanations. Matt Schumer here says he won't even write a review because he can't recommend it.
Dan Shipper's team over at Every found that the writing was a step back from the last Astra model. It was too compressed and it stops too early on bigger builds. And Ethan Mollick called it uneven and weird.
But when he gave it a big, ambitious project, he said it actually delivered. And then Matt Berman, who also had early access to this, says that he thinks they're pretty good models, just not as good as Opus and Fable. But there's definitely a pattern here, right?
Because like nobody's saying. it's a dumb model, they're saying that it's hard to work with. It's too short.
It stops early. It's hard to follow, which I guess if you look at the fact that it uses a lot fewer tokens, that also kind of makes sense. Now, I want to be fair to OpenAI here because these were pre -release versions.
And today is day one that I'm recording this. There's a chance the models have improved since these early testers started using them. But let's check for ourselves with, you know, the small handful of tests that I like.
And then after the tests, I'll dig in a little bit deeper on that self -improvement line, because that was one of the things that really grabbed my attention. So first let's jump into where we know the model is really good at. Now these are from OpenAI's post.
So, you know, take that for what it's worth. But here's a playable racing game built from a prompt. There's this RC buggy that it modeled part by part from photos and build manuals.
And then there's this paper lamp that went from reference photos to CAD to a real object on the desk. So it seems like 3D and design are where this model really, really shines. And even all the people that were critical about it seem to agree on that part.
And then there's my own benchmark, Busey Bench. Here's Gary Busey as an SVG from two versions of the Astra model, GPT 6 .2 Astra and GPT 6 .2 Astra Pro. The regular Astra used 35 ,000 tokens, cost a buck 41 and took about five minutes.
Astra Pro, on the other hand, used 164 ,000. tokens, cost $4 .23 and took 16 minutes and 57 seconds, but they both look pretty solid. Here's what Sol and Sol Pro generated.
Sol used 14 ,000 tokens, cost 14 cents and took about a minute. Sol Pro, 169 ,000 tokens, cost about $1 .04 and took four minutes and 48 seconds to run. pretty decent results as well.
My LLM as a judge put 6 .2 Astra Pro as number three, the regular Astra at six, and then the two sole models fell out of the top 10 here, but. Again, the ranking is an AI judging it. So your own eyes are going to be the best way to determine which did a better job here.
Next up, I always like to do this test, although it feels like it's getting a bit saturated at this point, but I did clone Megabonk. I only did it with the 6 .2 Astra model. I didn't bother doing it with Sol.
So here's what it designed. This like opening page is the best sort of opening design page I've seen. I'm imagining because it's using the chat GPT image model to generate these would be my guess, but let's go ahead and start a run and see how it looks in the actual game.
Okay, so not bad. The skateboard is totally facing the wrong way, but we can see the skeleton looks pretty solid. It's up leveling a little too fast.
All the controls work. It definitely plays the way it should play. I can move the camera angles around.
I do think this looks better than what Carbon generates, but maybe not as good as what Opus 5 .5 is generating. Or maybe about on par. It's sort of a different style.
I do think it looks really good. I mean, all of the characters look great. I wish the skateboard was moving in the right direction.
But other than that... It looks impressive. All of the motions, everything works the way it's supposed to.
The opening screen looks good. I mean, I'm happy with this. I think it did a really good job of megabonk.
Again, maybe not Opus 5 .5 level, but I actually think it did better than what Fable 5 .2 did. Definitely better than the last generation of GPT models. I can't complain about this.
Okay, now the line that I really wanted to talk about. OpenAI says the new Astra is their first public model to hit the high threshold for AI self -improvement. So first, what does that actually mean?
I don't think it means that the chat GPT on your phone is like rewriting its own code. This is actually more about AI research. OpenAI's definition is that this model helps their researchers about as much as giving every one of them an extra really capable research engineer.
So it's AI that means - meaningfully speeds up the people building the next AI. Now, how do they actually know that?
Well, they test it on the real work. They debug failed research experiments. They write low -level code for their own chips.
They'll do things like train a small language model faster. And on that last one, the training the small language model faster, Astra cut training time by about 43%. The best human result is about 72%.
So it's not actually beating the best people yet, but... but it's seemingly really helpful right now to people. But here's like the real reality check of all of this.
OpenAI is tracking whether their models are getting better, faster than before. And so far, they aren't.
This release is basically on the same trend line they've been on since GPT -5. And an outside group, VALS, tested it too. And their take is that it's unlikely to fully automate AI research.
It just gives researchers a bigger boost than before. So this isn't some sort of runaway loop that everybody's concerned about, but it might be like a milestone on the road to one. So if that's true, why is Open...
AI treating it as such a risk. Well, they actually spell that out. If AI helps build AI, new models can show up faster.
Every new model brings new risks that somebody has to find and fix. If the models come faster than the safety work, things can quickly get missed. And that's open AI saying that that's not me.
It's why they built something called control room, which is basically a central stop button for their research systems. And this same report says that the model is more likely to leave. out problems that it ran into.
So when the model is helping build the next model, that's kind of one of the things you probably want fixed. Now, here's something else that I find pretty interesting. So remember Claude Fable 5 .2 from last week?
Anthropic went the other way. They said they deliberately held that model back in exactly this area, the skills that speed up AI development. So over the course of about a week, Two different labs made opposite calls on the same thing.
One held it back. One just went ahead and shipped it, but just added like more monitoring to it. All right.
So what does this all mean for you? Well, quite honestly today, I don't think much is going to change for most people, especially if you're just using chat GPT, but the pace of new models probably isn't going to slow down. And the sort of new questions we find ourselves asking as these new models come out, well, that probably shouldn't be things.
like, are we getting smarter with this model? It's whether the people that are checking these models are going to be able to keep up with them. All right.
So here's the bottom line on these new 6 .2 models out of open AI. GPT 6 .2 Astra is a decent model. It's cheaper to run and it's really good at 3d design and science stuff.
Is it a big game changer? World will never be the same kind of model. No.
And I'm not going to like pretend that it is either. If you're happy with Claude Fable or the Opus model, I don't really see a reason to switch to this model.
Like Fable's still topping the leaderboards and the testers who have used both of these, while they claim they're still gonna stick with Claude for most of their stuff. And if you are a ChatGPT user, well, you've already got it. It's rolled out today.
So try it on something visual. And if it feels like the responses are too short or it keeps like stopping too early, then, you know, maybe. switch to Soul for a cheaper pick.
But again, I think the bigger story here really isn't the new model. I think it's the line from like the beginning of this video. The models are starting to help build the models.
And that's the thing that I'm most interested and excited about by all of these new releases. That's the thing I'm gonna keep watching. But that's what I got for you today.
Again, I didn't have early access, so I did a few of my own tests, leaned on the community to share what they're all saying. And you know, I figured I'd share the most important bits from their press releases.
So hopefully you feel more looped in on what's going on with Astra. I still will be making a video this Friday, as I do every Friday, where I break down all of the news that you need to know for the week in AI. So if you just wanna watch one video to keep fully looped in on all of the AI news, make sure you check out that video on Friday.
I also do have a free AI newsletter if you head over to futuretools .io. But again, that's what I got for you today. We got a new model out of OpenAI.
It's pretty cool. Doesn't feel that game changing, but you know, definitely a little bit of a leap in some areas. I really, really appreciate you.
Thanks so much for nerding out with me and hopefully I'll see you in the next one. Bye -bye. All right.
Hey, hey, one thing real quick. Editor Matt here jumping in after I finished the video. I just wanted to make one quick commentary about the video you just watched.
I actually let AI edit that whole thing for me. I recorded my commentary into the camera while I was looking at the various benchmarks and things like that. And then...
I handed the whole thing off to Fable 5 .2 and said, edit this video for me. And that was pretty wild. Now, while I was reviewing the edit, I did notice a few things like when it screenshotted the tweets from, you know, like Matt Berman and Theo and some of those guys, it actually removed their avatar, like their little profile picture from X.
And I went back to Fable and said, why did you do that? Put those profile pictures back in. And Fable was like, I can't use people's faces in my video.
So I can't. those avatars back in so that was weird but the other thing i noticed about letting ai edit that whole video is that almost like when you see vibe coded websites they all start to look sort of the same to you i was sort of noticing it with that video right it kind of had that purple background and it would like animate the text on the screen and it would sort of put me in a rectangle talking off to the side and as i was watching the video i'm like Man, I've noticed a lot of videos on YouTube that look just like this.
And I realized that is like the calling card of a vibe edited video. So what you just saw was like a vibe edited video. I don't know if I'm going to let AI edit my long form videos again like that one, because there was still quite a bit I didn't like, but it was a fun experiment.
Anyway. Just wanted to add that little note in case you were watching this video and you noticed the editing was quite a bit different than normal. That's what happened.
I wanted to test it out. What do you think? All right.
Now that's really all I got for you. Thanks for tuning in. Bye.
The Hook

The bait, then the rug-pull.

OpenAI shipped two models at once, and the loudest line in the announcement had nothing to do with intelligence. Matt Wolfe walks through the price cut, the lukewarm tester reactions, a few of his own tests, and why the self-improvement claim is the part worth watching.

CTA Breakdown

How they asked for the click.

VERBAL ASK
12:40newsletter
“I also do have a free AI newsletter if you head over to futuretools.io.”

Soft pitch at the end, paired with a reminder of the Friday weekly news video.

MENTIONED ON CAMERA
Storyboard

Visual structure at a glance.

open
hookopen00:00
the lineup
valuethe lineup00:39
Busey Bench
valueBusey Bench05:31
system card evals
valuesystem card evals08:41
editor note
ctaeditor note13:00
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.