OpenAI's GPT-6.2 Astra undercuts rivals on price, but early testers hit snags
Astra and the cheaper Sol model arrive a week after new releases from Anthropic and Google. An OpenAI lead says some reviewers may have missed last-minute changes.

Key Takeaways
- OpenAI says GPT-6.2 Astra sets new benchmark highs at less than a third of its predecessor's cost per task.
- Astra's API rates drop 20% from GPT-6 Astra, while Sol keeps GPT-6.1 pricing.
- The $200 Pro plan gets Ultrafast mode for the first time, starting with Sol.
- Some early testers reported models stuck in loops and abandoning tasks during early access.
OpenAI on Nov. 17 released GPT-6.2 Astra and GPT-6.2 Sol, two models it is selling mainly on price rather than raw power. The company says Astra posts new benchmark highs on professional and scientific work while costing less than a third as much per task as the Astra it replaces.
Astra is now OpenAI's most capable model for complex professional work, scientific research and engineering. Sol is the cheaper option for everyday writing, coding and multistep assignments. Both are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, and developers can reach them through the API.
The release came a week after Anthropic launched Claude Fable 5.2 and Mythos 5.2 and Google introduced Gemini 4 Hydrogen. Anthropic said Fable 5.2 beat GPT-6 Astra, Google's Gemini 4 Carbon and its own Opus 5.5 on nearly every benchmark it published, including coding, cybersecurity and knowledge work. In 9to5Google's reading of OpenAI's charts, the new models roughly match Anthropic's latest while often costing far less to run.
Competing on cost is a sensible choice. When the major labs release new models within days of one another, a benchmark lead rarely survives the next launch. A lower cost per finished task is harder to erase, and it matters more as products like ChatGPT Work take on whole projects. An agent that writes, tests and revises code across many steps uses far more tokens than a chatbot answering a single question.
What Astra and Sol cost
For developers, Astra costs $8 per million input tokens, 40 cents per million cached input tokens and $40 per million output tokens. OpenAI says its standard input and output rates are 20% lower than GPT-6 Astra's. Sol keeps GPT-6.1 Sol's rates of $2 for input, 10 cents for cached input and $10 for output, all per million tokens.
A 20% price cut alone can't explain a cost per task below one-third of the old figure. Most of the claimed savings must therefore come from the models finishing work with fewer tokens, and OpenAI says both models complete demanding tasks with fewer output tokens. Subscribers on flat-rate plans won't see a lower bill. The company says the gain shows up as more work done before they reach their usage limits.
Sol may be the better deal for many developers. On Cognition's FrontierCode 2.0, which tests whether changes to real codebases are good enough for maintainers to merge, OpenAI says Sol scored within half a point of Astra at roughly one-fifth the cost, with both models at maximum reasoning effort. If independent testing confirms that result, Sol is the obvious default for routine coding, and Astra becomes something teams switch to only for their hardest problems. OpenAI calls Sol "smarter, more efficient, and lets you complete more tasks on the same paid plan."
Templates, cleaner code and Ultrafast mode
OpenAI's examples for Astra focus on work people actually hand in. Give it a screenshot of one slide, and the company says Astra can use it as a template for the rest of the deck, matching the layout, colors and font size. OpenAI also says Astra's documents, spreadsheets and slides stick more closely to the templates users supply and the conventions of their field, and that it builds better-looking and better-working front ends for websites, games and 3D design.
In Codex, OpenAI says Astra is better at making and testing changes inside an existing project. "Early testers reported cleaner code style and more focused tests, with fewer rounds of revision," the company said.
The launch also extends Ultrafast mode, which OpenAI says can generate tokens up to 4.5 times faster than the standard rate on either model. Subscribers on the $200-a-month Pro plan get Ultrafast for the first time, starting with Sol. The $500-a-month Pro tier can use it with both models, as can eligible Enterprise and Edu workspaces once an administrator turns it on. Users switch speeds in the model picker.
Ultrafast draws down a user's allowance faster, so it works against the efficiency pitch behind the rest of the launch. Giving the $200 tier only Sol, and reserving fast Astra for the $500 tier, keeps OpenAI's most expensive model in its most expensive plan.
Early testers found rough edges
Some people who tried the models before launch were less impressed than the benchmarks suggest. Android Authority reported that a recurring complaint among early-access testers was that the models got stuck in unproductive loops, dropped assignments before finishing and wrote summaries that were harder to follow than necessary. In one case it described, several days of attempted coding improvements reportedly left a project running worse than when it started.
The same testers also had praise. According to Android Authority, they credited Astra with catching problems other models missed, finding unconventional solutions and producing strong 3D work. Their main frustration was getting that ability to show up reliably.
Thibault Sottiaux of OpenAI said on X that the company made substantial changes in the last three to four days before release and that he believed some reviewers missed them. He also acknowledged that OpenAI still plans to improve the models' creative taste and ease of use.
Late changes are a plausible explanation, and they mean the models customers use now may behave differently from the ones testers tried. They also mean early reviews of these models are of limited value, and the specific complaints still need to be checked against the released versions.
The complaints also cut against OpenAI's main pitch. Spending fewer tokens on a task only saves money if the task gets done. A model that loops, gives up early or makes a codebase worse can burn through more time and usage than a pricier model that finishes the job. For teams putting these models to work in ChatGPT Work and Codex, completion rates will matter more than benchmark scores, and the cost-per-task figures OpenAI is promoting depend on them.
The GPT-6 line has moved quickly. Astra debuted in September, Sol moved to version 6.1 a few weeks later at OpenAI's DevDay, and GPT-6.1 Terra followed. At that pace, OpenAI may have another chance to fix these issues within weeks. For now, Sottiaux's comments confirm that the company sees creative taste and ease of use as unfinished work.
Sources
- Android Authority — Android News: OpenAI launches GPT-6.2 Astra and Sol, but not everyone is sold on the cheaper upgrade
- 9to5Google — Google News: OpenAI releases GPT-6.2 Astra to regain performance lead from Claude Fable 5.2
- 9to5Mac — Apple News: OpenAI upgrades ChatGPT and Codex with new GPT-6.2 Astra and Sol models




