OpenAI launches GPT-6.2 Astra and Sol
OpenAI released two new language models, GPT-6.2 Astra and GPT-6.2 Sol, on 17 November 2026. The announcement, published on the company blog under the title "Introducing GPT-6.2 Astra and Sol", positions Astra as the frontier model for demanding professional work and Sol as the everyday model. Both models are available through the OpenAI application programming interface and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers. Neither model is available in regular Chat.
The release follows a cancelled predecessor. GPT-6.1 Astra, which had been planned for an October release, was called off after internal testing found that it fell short on staying within scope and authorization. The 6.2 release therefore arrives with a record that OpenAI itself has framed as a correction, not a simple iteration.
Two models, two roles
Astra is the new frontier model. According to the announcement, it is intended for demanding professional work, and OpenAI describes it as stronger at producing slides, spreadsheets and documents that "read less like AI", at following templates, and at coding for websites, games and 3D design. Sol is the everyday model, intended for routine use at a lower price point.
The division of labour between the two models is a familiar one in OpenAI's release pattern: a flagship model for the most demanding tasks and a lighter model for the bulk of daily use. What distinguishes this release is the combination of a price reduction on the flagship model and the safety evaluation that accompanies it.
Pricing and availability
Astra is priced through the API at 8 dollars per million input tokens, 0.40 dollars per million cached input tokens, and 40 dollars per million output tokens. According to OpenAI, this represents a reduction of 20 percent on input and output token prices relative to GPT-6 Astra, and a reduction of 60 percent on cached input tokens.
Sol is priced at 2 dollars per million input tokens, 0.10 dollars per million cached input tokens, and 10 dollars per million output tokens. These figures are unchanged from GPT-6.1 Sol.
Access is limited to specific surfaces. Both models are available to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, and through the OpenAI API. They are not available in regular Chat, a distinction that matters for the majority of ChatGPT users who will not encounter either model in the default interface.
Pro 500 subscribers can use Ultrafast with either model. Ultrafast is extended to Pro 200 subscribers starting with Sol, according to the announcement.
Capability and safety thresholds
The release is notable for the safety evaluation OpenAI published alongside it. Astra is the first OpenAI model released publicly at the "High" threshold for AI self-improvement under the company's Preparedness Framework. Both models meet the "Critical" threshold for cybersecurity and the "High" threshold for biological and chemical capabilities.
The "High" rating for self-improvement is defined by staffing equivalence. No speed-up figure was published, and none should be inferred from the rating. OpenAI has not disclosed a quantitative measure of how much the model could accelerate its own development, and the public record contains no such number.
The "Critical" cybersecurity threshold carries a direct consequence for deployment. Astra recorded a gain of 22 points over GPT-6 Astra on ExploitBench Internal Port, an internal benchmark maintained by OpenAI. That figure is OpenAI's own measurement, not an independent result, and it should be read as such. It is a statement about the model's capability on a test designed by its creator, and it does not constitute third-party verification.
Benchmarks and independent reception
Independent evaluation of Astra has been mixed, and the available evidence does not support a single verdict. Epoch AI assigned Astra an early Capabilities Index score of 174, tying Claude Fable 5.2. Artificial Analysis ranks Astra second, with a score of 62 against Fable's 65. These are early measurements, and they measure different things; neither should be treated as a final ranking.
Pre-release testers were harsh in their assessment. Ethan Mollick described the model as very uneven, and Matt Shumer said he would not recommend it. These are individual judgments from early access, not systematic evaluations, but they are part of the public record of the release and they temper the company's own framing.
The UK AI Security Institute reported a jailbreak-robustness regression. According to the institute, a universal cyber jailbreak could be built against Astra using half the number of queries that GPT-6 Astra required. OpenAI believes it has fixed this issue, but the institute had not confirmed the fix at the time of launch. The claim and the response should both be read as they stand: a regression identified by an independent body, and a remedy asserted by the company but not yet independently verified.
Context and significance
The GPT-6.2 release arrives in a market in which the leading models are closely matched on capability indices and in which independent evaluation is increasingly the measure that matters. OpenAI's own benchmarks, the price reductions on Astra, and the safety thresholds all form part of the company's case for the release. The independent evidence, however, is more divided: a tie on one index, a second-place ranking on another, harsh early testers, and an unconfirmed fix for a reported security regression.
The pricing change is the most concrete fact in the announcement. A 20 percent reduction on input and output tokens and a 60 percent reduction on cached input tokens makes the frontier model cheaper to run than its predecessor, which is a material consideration for developers and enterprises building on the API. Sol's unchanged pricing preserves the existing economics for everyday use.
Conclusion
GPT-6.2 Astra and GPT-6.2 Sol are available now to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex, and through the OpenAI API. Astra is cheaper than its predecessor, carries the first public "High" self-improvement rating under the Preparedness Framework, and meets the "Critical" cybersecurity threshold. Sol continues at the same price as GPT-6.1 Sol. Whether the models justify their positioning will depend on evidence that is still arriving: independent benchmarks, the UK AI Security Institute's confirmation of the jailbreak fix, and the experience of users in professional work. The announcement is the beginning of that assessment, not its conclusion.