OpenAI brings GPT-6 Sol and Luna to ChatGPT and adds an Intelligent UI for mixed-format answers

OpenAI announced on 7 October that it is bringing new GPT-6 models to ChatGPT, following the first GPT-6 models it introduced in September. According to 9to5Mac, subscribers on the Plus, Pro, Business and Enterprise plans are getting GPT-6 Sol first, and GPT-6 Luna is due to start reaching Free and Go users the next day. OpenAI describes these versions as tuned for everyday conversation. In the system card for the release, the company says GPT-6 is coming to free and paid plans worldwide and that the October models replace the GPT-5.6 versions of Sol and Luna in ChatGPT.

The most visible change is what OpenAI calls Intelligent UI. Instead of replying in text alone, ChatGPT can now mix text with visual and interactive elements, choosing the format to fit the question and what the answer contains. OpenAI lists graphics, buttons that can be tapped, forms, charts and interactive experiences used inside the chat. A comparison might be laid out side by side and an explanation might come as an interactive diagram, while a plain text reply remains an option when that serves the user best. For more complex prompts, ChatGPT can build widgets that let people explore a concept or change inputs and watch the results update, and it can create small tools such as calculators, bill splitters and games within a conversation.

OpenAI also says GPT-6 lets ChatGPT start responding before its reasoning is complete, adding information to the reply step by step rather than holding everything back until the thinking process ends. The update does not reach every OpenAI product yet. The system card says people using the models in Codex or through ChatGPT Work are still on earlier releases. OpenAI labels the models by release month: October for the versions now in ChatGPT, September for those that remain in Codex and Work.

The safety claims in the system card come from OpenAI's own testing. The company says the October models resist jailbreaks better than the GPT-5.6 models, including attacks that adapt over several turns, and show less dishonesty and deception. It says it updated its safety training to better guard against high-risk misuse involving cyber, biology and violence. OpenAI also reports statistically significant regressions in some content categories, such as self-harm and gore, and in several evaluations for users under 18; it says manual review and red-teaming found the failures were generally of low severity and that further safeguards, such as classifiers, operate in production. On its instruction hierarchy tests for prompt injection, OpenAI gives robustness scores of 99.99% for GPT-6 Sol and 99.79% for GPT-6 Luna.

OpenAI's Preparedness Framework assessment places both October models at High capability in the Cybersecurity domain and in the Biological and Chemical domain. The company says neither model reaches its High threshold for AI Self-Improvement and that both remain below the Critical level in cybersecurity. That matches its rating of the GPT-5.6 models, and OpenAI says it applies the same safeguards it described in the GPT-5.6 system card. In its own cyber evaluations, it reports that GPT-6 Sol landed roughly level with its predecessor with no clear gain, while GPT-6 Luna scored below GPT-6 Sol.