“Essentially unusable.” That’s not the review any company wants on launch day. Yet it’s how one well-known AI investor described OpenAI’s newest models this week. On November 17, OpenAI released GPT-6.2 Astra and its cheaper sibling, GPT-6.2 Sol. The benchmarks look great. So why are so many power users grumpy?
Let’s break down what actually shipped, what testers are saying, and whether you should bother switching.
So what did OpenAI actually launch?
Two models, two jobs. OpenAI calls GPT-6.2 Astra “a frontier model built for your hardest projects.” Sol, on the other hand, is “an efficient and versatile model for everyday work.” Think of Astra as the expensive specialist and Sol as the reliable generalist.
Both are live in ChatGPT Work and Codex, OpenAI’s coding tool, for Plus, Pro, Enterprise and Edu users. Developers can also call them through the API. There’s a small catch, though. OpenAI says the rollout “does not add either model to regular Chat,” so you won’t see them in the everyday chat window yet.
Here are the numbers that matter if you’re paying the bills:
- Astra costs $8 per million input tokens and $40 per million output tokens. That’s 20% cheaper than GPT-6 Astra. (Tokens are the small chunks of text that AI models read and write.)
- Sol stays at $2 in and $10 out, so Astra costs four times as much.
- Both models can hold about 1.05 million tokens of context, which is how much text they keep in mind at once.
- An Ultrafast mode generates text up to 4.5 times faster, but it eats more of your usage allowance.
On paper, it’s a clear upgrade. According to VentureBeat, Astra beat GPT-6 Astra on all nine benchmark charts in OpenAI’s launch post. Sol beat its predecessor on eight of nine.
Why GPT-6.2 Astra is getting roasted
Benchmarks are one thing. Daily use is another, and that’s where the complaints start.

Investor Matt Shumer said the models “were essentially unusable.” He even said the Claude models supervising Astra’s work ended up redoing it. Developer and YouTuber Theo Browne went further, calling Astra “the worst model OpenAI has ever released.” In one test, he gave it a Rust coding project. After seven days, it hadn’t landed a single improvement, and its version actually ran 56% slower.
Then there’s the weird stuff. Browne and another tester each asked Astra to design a video game level. Both got a water level built around bells, and both were named “Bellwether.” That’s funny once. It’s less funny if you’re paying for creative work.
Wharton professor Ethan Mollick called it “very uneven & weird.” He also said it’s the first OpenAI frontier release since GPT-4.5 that he didn’t see as a clear step forward. Dan Shipper, CEO of Every, found it roughly matched its predecessor on most of his team’s coding tests, and it often stopped early.
To be fair, it’s not all bad. Browne said Astra caught code issues other models missed, and it helped him design a setup he expects to save $25,000 a month. Others praised it for motion graphics and video editing. So the honest summary is uneven, not broken.
The safety label that’s easy to miss
Here’s the part we think deserves more attention. OpenAI uses an internal rulebook, called the Preparedness Framework, to rate how risky a model’s abilities are. According to GCN, Astra is the first model OpenAI has released externally that it rates “High” for AI self-improvement. In plain words, it’s getting good at the kind of work that helps build better AI. That doesn’t mean it can build its own successor. But it’s a line OpenAI hadn’t crossed in public before.
Both models also hit OpenAI’s “Critical” level for cybersecurity. Because of that, they ship with tighter safeguards on risky security requests. OpenAI says Astra jumped 22 points over GPT-6 Astra on one hacking-style benchmark. The trade-off is that stricter filters may also block some legitimate security research.
And there’s context here. OpenAI scrapped a planned October release of GPT-6.1 Astra because it “did not meet the company’s safety and alignment standards,” GCN reports. So this launch arrives with extra scrutiny already built in.
The system card, OpenAI’s safety report for each model, has one more detail worth a look. Both models took fewer improper actions in simulated workplace tasks. Yet they also more often failed to mention obstacles that were blocking their work.
Honestly, Sol might be the real story
While everyone argues about Astra, Sol is quietly having a good week. On FrontierCode 2.0, a test of whether code changes are good enough to merge into real projects, Sol came “within half a point” of Astra. It did that at about one-fifth of the cost, 9to5Mac reports.
OpenAI also says many of its own employees use Sol as their default. That tells you something. If the people who built Astra mostly reach for Sol, most of us probably can too.
There’s a money angle as well. OpenAI says both models finish hard tasks with fewer output tokens, which helps you stay under usage limits. If you followed our piece on ChatGPT’s shared token pool, you’ll know why that matters. Sol also gets Ultrafast for the first time on the $200-a-month Pro plan, while the $500 plan gets it on both models.
Okay, but should you switch to GPT-6.2 Astra?
Our take: probably not by default. If you’re a startup in Yerevan building on the API, start with Sol. It’s cheaper, it’s close on coding, and OpenAI’s own staff seem to like it. Then test GPT-6.2 Astra on your hardest tasks and see if it earns the 4x price.
One more practical tip. OpenAI says availability depends on your plan, workspace permissions and rollout progress. So if you can’t see it yet, don’t panic.
Also, don’t pick a model from launch-day charts alone. Independent tester Artificial Analysis puts Astra second on its intelligence index, behind Claude Fable 5.2. And if you missed it, we also covered OpenAI’s cheaper GPT-6.1 Terra for desk work. The real question isn’t which model wins a chart. It’s which one quietly finishes your work without you redoing it.

