All guides
    AI news
    4 min read

    OpenAI releases GPT-6.2 Astra and Sol

    OpenAI's new frontier model, GPT-6.2 Astra, is the first it has released that it rates High for AI self-improvement. A cheaper model, Sol, launched alongside it.

    OpenAI releases GPT-6.2 Astra and Sol

    On November 17, 2026, OpenAI released two models: GPT-6.2 Astra and GPT-6.2 Sol. Decrypt, Digital Trends, Implicator and OpenAI's own announcement all cover the launch. Astra is described as the frontier model for demanding work. Sol is the cheaper, everyday model.

    What was reported

    • Both models are available in ChatGPT Work, Codex and the API, according to coverage of the launch.
    • Astra's API price is reported as 20% below its predecessor, GPT-6 Astra, at $8 per million input tokens and $40 per million output tokens. Sol is reported to keep the prices of GPT-6.1 Sol.
    • OpenAI's system card rates both models Critical for cybersecurity and High for biological and chemical capabilities, as reported.
    • Astra is the first model OpenAI has released publicly that it says meets its High threshold for AI self-improvement. Sol falls below that threshold, per the system card as reported.
    • The independent benchmark firm Artificial Analysis found no improvement on its Coding Agent Index: Astra was flat against GPT-6 Astra, and Sol fell three points against GPT-6.1 Sol.

    We were not able to open OpenAI's own pages from our research setup, so the figures above come from press coverage of the announcement and system card. Check them against OpenAI's release post before relying on them.

    What the self-improvement rating means

    OpenAI's safety framework sorts models by how much they could help with certain risks. The self-improvement category measures how much a model could speed up AI research itself. Reaching High does not mean a model is improving itself without people. It means the company has judged that the model crosses a line it set in advance, which triggers extra safeguards and review. One outlet noted that, on a separate capabilities index, the card says Astra is still within the existing trend, and OpenAI says it continues to evaluate that capability as part of its safety work.

    The Critical rating for cybersecurity is the other detail worth noting. It is a statement about what a model could be misused to do, not a report that anything has happened. We will not describe how such misuse works.

    Why it matters

    The pace of releases is itself the story. A frontier model that arrives with a price cut, together with a cheaper sibling, makes capable AI easier for companies and developers to build into everyday products, including ones used by children and schools. The independent benchmark result is a useful counterweight: a new version number does not always mean a better result for every task, and the coding scores here were flat or lower.

    It is also a reminder that the companies publish their own risk ratings, and that the ratings come from the same companies selling the models. Outside testers such as Artificial Analysis help, but they measure performance, not safety.

    What it means for families, schools and frum professionals

    • Parents: the model behind a chat tool can change without any announcement to the user. Ask any AI product your child uses what filtering it adds on top of the model, because the model's ratings do not describe the product's safeguards.
    • Schools and mechanchim: cheaper, stronger models make AI easier to add to existing software. A school policy should name which tools are approved rather than assuming a vendor's model is unchanged from last term.
    • Frum professionals: coverage says OpenAI is pitching the new models for finance, law and accounting work. Confidentiality of client information still depends on the contract and settings of the specific product, not on the model's version.
    • Halachic questions: AI never paskens. A question about practice belongs with a rav.

    What it changes for someone using Navōn

    Nothing directly. This is a story about another company's models. Navōn's safety check runs on the server, where it applies to what comes in and what goes out regardless of which model is answering. A new model release does not change that principle, and no AI tool, including Navōn, should be treated as perfectly safe.

    See what Navōn's filtering and parent controls do today.

    Common questions

    What is GPT-6.2 Astra?

    OpenAI's frontier model, released November 17, 2026 for demanding work. It is available in ChatGPT Work, Codex and the API, according to coverage of the launch.

    What is GPT-6.2 Sol?

    A cheaper, everyday model released alongside Astra. Coverage says it keeps the prices of GPT-6.1 Sol.

    What does High for AI self-improvement mean?

    It is a rating in OpenAI's own safety framework for how much a model could speed up AI research. It does not mean the model improves itself without people.

    Is the new OpenAI model safe for children?

    A model's risk ratings do not describe how safe a product built on it is. Ask what filtering the product adds, and keep adult supervision in place.

    Sources

    Published by Navōn. How these guides are written and checked.

    Read next