OpenAI's GPT-6 arrived in stages. It was announced in early September with GPT-6 Astra, and on October 7 OpenAI began launching GPT-6 to everyone in ChatGPT, across free and paid plans worldwide. GPT-6 Sol and GPT-6 Luna replace the GPT-5.6 models that ChatGPT used before.
The launch itself is a big moment. The document OpenAI published with it is, in some ways, the more interesting read.
A Frontier Model on the Free Plan
When the newest model reaches free users on day one of a general rollout, expectations move for everyone. People who use ChatGPT at home will carry its quality bar to every product with a text box and an AI label — including yours. A feature that felt impressive in spring can feel dated by autumn without anything in it changing.
What OpenAI Disclosed
OpenAI's deployment page for the October update is unusually candid. Alongside the improvements — stronger resistance to jailbreaks, including multi-turn attacks, and fewer deceptive or guardrail-circumventing behaviours than GPT-5.6 — it lists where the new models do worse on its own safety evaluations, including some self-harm, sexual-content and under-18 evaluations. OpenAI notes that the failures it reviewed were generally low severity, and that some system-level protections are not reflected in those scores.
Under its Preparedness Framework, OpenAI treats both models as High capability in cybersecurity and in biological and chemical domains, and applies the safeguards from the GPT-5.6 system card. Neither model reaches the High threshold for AI self-improvement.
Publishing regressions next to improvements is the right practice, and it is useful to anyone building on these models: it says exactly where to look first in your own testing.
New Models Are Not Drop-In Upgrades
A model that is better on average can still be worse at the specific thing your product relies on. The October page itself shows this — the same release improves some evaluations and regresses others. For teams with AI features in production:
- Pin model versions in the API and upgrade on purpose, not because a default changed.
- Keep an evaluation set of real prompts with expected behaviour, and run it before every model switch.
- Read the vendor's safety and system documents for areas that touch your users, such as minors, health or security.
- Keep your own guardrails. Validation, rate limits and fallbacks in your application do not depend on any one model behaving well.
The Takeaway
Model releases now come every few months, and each one moves both what is possible and what users expect. The teams that benefit are the ones that can test a new model against their own product in a day — and decide, with evidence, whether to switch.