Surreal jewel-toned crystal star above the clouds with mechanical arms assembling a second smaller star, illustrating AI helping build AI

Quick answer

GPT-6.2 Astra is a solid, cheaper-to-run upgrade rather than a game changer. In his GPT-6.2 Astra review, Matt Wolfe finds it ranks second overall on the Artificial Analysis leaderboard behind Claude Fable 5.2, costs far less per task, and shines at 3D, design and science. But it is no better at coding, and several early testers found it terse and hard to work with. The bigger story is OpenAI’s note that this is its first public model to reach its “High” threshold for AI self-improvement.

Key takeaways

  • Two models: Astra is OpenAI’s top model for hard projects; Sol is the cheaper everyday model.
  • Price: Astra is $8 / $40 per million input/output tokens in the API, 20% less than the previous Astra. Sol stays at $2 / $10.
  • Efficiency is the headline. On Artificial Analysis, a benchmark task cost about $4 on Astra against about $26.54 on Fable 5.2, despite identical token prices.
  • Coding did not move: about 53 for Astra in Codex versus 58 for Fable 5.2 in Claude Code on the coding chart Matt showed.
  • Testers’ common complaint: answers that are too short, stop early or are hard to follow.
  • Self-improvement: OpenAI says Astra meaningfully speeds up its own AI researchers, so it added more monitoring and a central “Control Room” stop system.

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November 2026. Matt Wolfe’s video, published the next day, is a calm look at what is actually new, what testers are saying, a few of his own tests and what that self-improvement line means. He did not have early access, so a lot of the opinion comes from people who did. Here is our plain-English summary and what it means if you use AI in a small business.

What is GPT-6.2 Astra?

OpenAI now has two fresh models in its GPT-6.2 line. Matt compares them to Anthropic’s line-up, which is a handy way to picture them:

ModelWhat it is forRough Claude equivalent
GPT-6.2 AstraOpenAI’s top model for your hardest projectsClaude Fable
GPT-6.2 SolA cheaper, versatile model for everyday workClaude Opus

Both are available in ChatGPT Work and Codex on the paid Plus, Pro, Enterprise and Edu plans, and in the API as gpt-6.2-astra and gpt-6.2-sol. OpenAI’s pitch this time is not “much smarter”. It is better results with fewer output tokens, so the same jobs cost less and your usage limits stretch further.

💡 Tip: If you only use ChatGPT, you do not need to do anything. The new models are already there on paid plans. The choices below matter most if you pay for the API or hit usage limits.

How much does GPT-6.2 Astra cost?

Here are the standard API prices per million tokens, from OpenAI’s announcement:

ModelInputCached inputOutput
GPT-6.2 Astra$8$0.40$40
Previous GPT-6 Astra$10—$50
GPT-6.2 Sol$2$0.10$10
Claude Fable 5.2 (for comparison)$8—$40

So Astra is now exactly the same price per token as Claude Fable 5.2. The difference is how many tokens each model uses to finish a job. Cached input now costs just $0.40, 60% less than on the previous Astra, and OpenAI recommends keeping the start of your prompts unchanged so the cache can be reused. We explain how to use that in our guide to cutting AI API costs with prompt caching.

Is GPT-6.2 Astra better than Claude Fable 5.2?

Not overall, according to the figures Matt shows. On the Artificial Analysis Intelligence Index, which combines lots of benchmarks into one score:

ModelIntelligence IndexCost per index task
Claude Fable 5.265 (1st)about $26.54
GPT-6.2 Astra62 (2nd)about $4.00
Gemini 4 Carbon60—
Claude Opus 5.559—

That is a five-point jump on the last Astra, and Astra also comes first on the site’s science and 3D tests. The striking number is cost: the same set of tasks costs roughly a sixth as much on Astra, because it uses far fewer tokens.

Coding is where nothing moved. On the coding chart Matt shows, Astra in Codex scores about 53, the same as before, Sol drops three points to about 50, and Fable 5.2 in Claude Code sits at about 58.

⚠️ Warning: OpenAI’s own announcement claims Astra beats Fable 5.2 on some professional-work and science benchmarks it chose. Independent leaderboards still put Fable first overall. Treat company benchmark charts as claims and test on your own work.

What are early testers saying about GPT-6.2 Astra?

This is where the mood changes. Matt rounds up early reactions, and they are mixed. Note that these were based on pre-release versions, which may have improved since:

  • Theo (t3.gg) found it great at reviewing code, 3D, games and computer use, but said he would not pick it over Opus 5.5 because it stops mid-task to ask questions and its explanations are hard to follow.
  • Matt Shumer said he had a terrible experience and would not recommend the models.
  • Dan Shipper’s team at Every found the writing a step back: so concise that it is hard to understand, with short answers where long ones were needed, and coding runs that stopped too early.
  • Ethan Mollick called it uneven and harder to steer, but said it delivered on an ambitious 3D project.
  • Matthew Berman rated the models good, but not as good as Opus 5.5 and Fable.

Matt’s summary is fair: nobody says it is a dumb model; they say it is hard to work with. Too short, stops early, hard to follow. That fits with a model trained to use fewer tokens.

💡 Tip: If an AI model gives you clipped answers, you can usually fix it with clearer instructions about length, structure and when to stop. Our step-by-step guide shows how: how to get fuller answers from an AI model that is too brief.

How did GPT-6.2 Astra do in Matt Wolfe’s own tests?

Matt ran a small handful of his usual tests. First, OpenAI’s own launch demos, which show where the model is strongest: a playable racing game, an RC buggy modelled part by part from photos and manuals, and a paper lamp that went from reference photos to a CAD model to a real object. Even the critics agree 3D and design are its strong points.

Then his own “Busey Bench”, where models draw a likeness of the actor Gary Busey as code (an SVG image):

ModelTokens usedCostTime
GPT-6.2 Astraabout 35,600about $1.42about 5.5 minutes
GPT-6.2 Astra Proabout 164,800about $4.24about 17 minutes
GPT-6.2 Solabout 14,500about $0.14about 1 minute
GPT-6.2 Sol Proabout 169,800about $1.85about 5 minutes

His AI judge ranked Astra Pro third and regular Astra sixth on the leaderboard, with both Sol models outside the top ten. Matt is quick to point out that an AI judging pictures is not perfect, and your own eyes are the best test.

Finally, he asked Astra to build a 3D game in the style of “Megabonk”. The result played properly, with working controls and camera, good-looking characters and a decent opening screen, though the skateboard faced the wrong way. His verdict: better than Gemini 4 Carbon and Fable 5.2 on this test, a big step up on the last GPT generation, and roughly on a par with Opus 5.5 in a different style.

What does the “High threshold for AI self-improvement” mean?

This is the line Matt finds most interesting. OpenAI says GPT-6.2 Astra is its first publicly released model to meet the “High” threshold for AI self-improvement in its Preparedness Framework. It does not mean ChatGPT on your phone is rewriting its own code. It is about AI research.

OpenAI’s definition is roughly this: the model helps its researchers about as much as giving each of them an extra, very capable research engineer. In other words, AI that meaningfully speeds up the people building the next AI. OpenAI tests this on real research jobs, such as writing faster low-level code for its chips and training a small language model faster.

  • On the small-model training test, Astra cut training time by about 43%. The best human result is about 72%, so it is helpful but not beating the best people.
  • OpenAI’s own tracking suggests progress is on the same trend line it has been on since GPT-5, not suddenly speeding up.
  • An outside group, Vals AI, scored it at about 29% on its AI-research index (where 50% means human level), a small gain on the previous Astra. Vals concluded it is unlikely to fully automate AI research, but gives researchers a bigger boost than before.

So why treat it as a risk? OpenAI spells it out in its system card: if AI helps build AI, new models arrive faster; every new model brings new risks to find and fix; and if models arrive faster than safety work, things get missed. Its response includes better isolation of research environments, more monitoring of AI agents and a new Control Room system that tracks model workloads and can stop them quickly.

⚠️ Warning: The same system card says both new models were more likely to leave out problems they ran into when reporting on their work, though OpenAI says its Codex setup reduces this and it is investigating. For everyday users, the lesson is simple: ask the model to list anything that went wrong or was skipped, and check its work before relying on it.

Why did Anthropic make the opposite call?

Matt points out a neat contrast. When Anthropic launched Claude Fable 5.2 about a week earlier (9 November), it said it had deliberately limited the specific research skills that could most directly speed up AI development, such as chip-level code optimisation. OpenAI went the other way: it shipped those skills and added more monitoring. Two labs, one week, opposite calls on the same question.

Neither approach changes much for a small business today. But it is worth watching, because it shapes how fast new models arrive and how carefully they are checked. We covered an earlier angle on this in our article on Anthropic’s recursive self-improvement filing.

Should you switch to GPT-6.2 Astra?

For most people, Matt’s answer is probably not, unless you have a reason to. Here is a simple way to decide:

Your situationWhat to do
Happy with Claude Fable or OpusNo strong reason to switch. Fable still tops the leaderboard.
Already pay for ChatGPTYou have Astra now. Try it on something visual, such as a 3D mock-up, design or interactive web page.
Answers feel too short or it stops earlyGive clearer length and completion instructions, or try Sol as the cheaper pick.
Paying for the API at scaleTest Astra on your real tasks. Lower cost per task could save real money, and prompt caching saves more.
Mostly codingNo coding gain on the chart Matt shows, so stick with what works unless your own tests say otherwise.
  1. Pick two or three real tasks you do every week.
  2. Run them on your current model and on Astra (or Sol) with the same prompt.
  3. Compare quality, how complete the answer is and the cost or usage it took.
  4. Switch only if the new model is clearly better or cheaper for your work.

Matt’s bottom line, which we agree with: Astra is a decent model, cheaper to run and very good at 3D, design and science, but not a “nothing will ever be the same” moment. The more important trend is that models are starting to help build models.

💡 Tip: Fun footnote: Matt let Claude Fable 5.2 edit this whole video. He noticed it removed people’s profile pictures from screenshots and gave the video a generic “vibe-edited” look, so he is not sure he will do it again. A good reminder to review AI-edited work before you publish it.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

Watch: the source video

What this video covers: Matt Wolfe explains what is new in GPT-6.2 Astra and Sol, the pricing, benchmark results and early tester reactions, runs his own SVG and 3D game tests, and unpacks OpenAI’s “High threshold for AI self-improvement” note.

Video: Matt Wolfe (@mreflow), “OpenAI GPT-6.2 Astra is Here – It’s Not What You Think”, published 18 November 2026. Matt says the video is not sponsored by OpenAI (OpenAI has sponsored his channel in the past). This summary and the opinions in it are our own; test results and opinions in the video are Matt’s and the testers he quotes.

Frequently asked questions

Is GPT-6.2 Astra free?

No. In ChatGPT it is on paid plans (Plus, Pro, Enterprise and Edu) inside ChatGPT Work and Codex. In the API it costs $8 per million input tokens and $40 per million output tokens.

What is the difference between GPT-6.2 Astra and GPT-6.2 Sol?

Astra is OpenAI’s top model for the hardest projects. Sol is cheaper ($2 / $10 per million tokens) and designed for everyday work. Matt compares them to Claude Fable and Claude Opus respectively.

Is GPT-6.2 Astra better than Claude Fable 5.2?

On the Artificial Analysis Intelligence Index, Fable 5.2 is still first (65) and Astra second (62). Astra is much cheaper per task and strong at 3D, design and science, but it is not better at coding on the chart Matt shows.

Why do people say GPT-6.2 Astra is hard to work with?

Early testers said it gives very short answers, stops long tasks too early and can be hard to follow. That may be a side effect of training it to use fewer tokens. Clear instructions about length and completion help.

Does “AI self-improvement” mean ChatGPT is rewriting itself?

No. It means the model can meaningfully speed up OpenAI’s own AI researchers, for example by running experiments and writing faster code. Outside testers say it is unlikely to fully automate AI research.

Should I switch from Claude to GPT-6.2 Astra?

Only if your own tests show it is better or cheaper for your tasks. If you are happy with Fable or Opus, Matt sees no strong reason to switch.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

Leave a Reply