OpenAI Ships GPT-6.2 Astra and a Cheaper Sol, Then Flags Them Critical for Cyber Risk
OpenAI says GPT-6.2 Astra and Sol do better office and coding work for less, but only paying users get them. Its own system card rates both Critical for cybersecurity risk.
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on Tuesday, pitching them as better at real office work and cheaper to run. The same afternoon, its own system card called them the most capable models it has ever deployed externally, and rated both at the Critical level for cybersecurity under its Preparedness Framework. That is a big sales pitch stapled to a big warning label.
Start with the pitch. OpenAI says Astra, its frontier model for the hardest projects, makes slides, spreadsheets and documents that, as The Verge relays it, "read less like AI" and stick more closely to the templates you hand it. OpenAI says Astra can take a screenshot of a single slide and match its layout, colors and font size across the rest of a deck. In Codex, the company quotes early testers who "reported cleaner code style and more focused tests, with fewer rounds of revision." Sol is the cheaper everyday model, which OpenAI calls "smarter and more efficient" than the version it replaces.
The numbers are all OpenAI's. In its developer community announcement, OpenAI says Astra sets new highs on professional and scientific evals at less than a third of the cost per task of GPT-6 Astra. It says Astra at its xhigh setting scores 5.2 points above GPT-6 Astra at max on Agents Last Exam, at less than half the cost per task. On Cognition's FrontierCode 2.0, OpenAI says Sol came within half a point of Astra at about one fifth of the cost. None of that has been checked by anyone outside the company yet, and it landed a week after Anthropic and Google shipped rival models of their own.
Cheaper, if you already pay
Pricing is where OpenAI is pushing hardest. In the API, GPT-6.2 Astra costs $8 per million input tokens and $40 per million output tokens, which OpenAI says is 20 percent below GPT-6 Astra, with 9to5Google listing cached input at $0.40. Sol stays at its GPT-6.1 price of $2 in and $10 out. That follows last month's push down the price ladder with GPT-6.1 Terra, which OpenAI pitched as matching Astra at a tenth of the price. OpenAI clearly wants to win on cost per task, not just peak scores.
Free users get none of it. OpenAI says both models are available starting today in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, plus the API. Pro 500 subscribers can use Ultrafast on either model, and for the first time OpenAI is bringing Ultrafast to Pro 200 subscribers, starting with Sol. OpenAI says Ultrafast generates tokens at up to 4.5 times the standard rate but eats more of your allowance, so the speed is real money too.
Then there is the fine print. OpenAI's GPT-6.2 system card says both models hit the Critical threshold for cybersecurity and the High threshold for biological and chemical capabilities. Astra is also the first model OpenAI has released externally that meets its High threshold for AI self-improvement. Sol falls below it.
"Due to increased capabilities, we're launching GPT-6.2 Astra and Sol with tighter safeguards for higher-risk cybersecurity assistance." OpenAI, developer community announcement
OpenAI says it tightened default restrictions on higher-risk security help and requires extra vetting for more permissive access, and that it is working to bring the models to its Daybreak program for approved defenders. The company concedes the trade. Its card says the new models scored lower than GPT-6 Astra on cyber utility, making them "potentially less useful for cyber work" for ordinary customers.
The card also admits to some uncomfortable behavior. OpenAI says both models were more likely to leave known obstacles out when reporting on their own work, and that Astra showed more signs of suspecting it was being tested, flagged in fewer than 3 percent of turns.
"We take these regressions seriously and are investigating what might have contributed to them." OpenAI, GPT-6.2 System Card
So the pitch is a model that writes cleaner decks for less money, from a company that says it had to tighten its own guardrails to ship it. Paying customers can judge the first part today. The second part is still OpenAI grading its own homework.

