OpenAI has updated GPT-5.6 Sol, the model behind ChatGPT for Plus and Pro subscribers, and pushed a new model, GPT-5.6 Luna, out to everyone using the free tier.
The company is also removing the rate limit on text conversations for free users, allowing them to keep chats going without waiting for the limit to reset.
For Plus and Pro accounts, GPT-5.6 Sol now handles both quick replies and longer reasoning through one model instead of switching between separate ones for “Instant” and “Thinking” modes.
OpenAI says the goal was a model that “delivers more focused answers, adapts its level of detail to the question, avoids unnecessary formatting” regardless of how much effort it puts into a reply. A new slider lets people set how much reasoning effort the model uses for a response.
OpenAI reports that the update cuts down on factual mistakes. In an internal test on financial, medical, and legal prompts, answers containing at least one factual error were about 62 percent less common with Luna and 68 percent less common with Sol compared with the outgoing GPT-5.5 Instant model.
What changed for free users
Free and Go tier users are being moved to GPT-5.6 Luna as their default model this week. Starting next week, OpenAI will remove the rate limit on text chats and add a new “Think” button for deeper reasoning on harder questions. Unlimited chats will remain subject to what OpenAI calls “abuse guardrails,” while limits remain in place for file uploads, images, and other tools.
Plus and Pro subscribers can already use the updated Sol model and its new slider. OpenAI notes that the version of Sol tuned for this update is limited to ChatGPT’s Chat experience. The version of Sol that powers Work and Codex is not part of this change.
OpenAI expands GPT-5.6 safeguards for younger users
According to the company, GPT-5.6 Sol was trained to avoid romantic roleplay, age-restricted challenges, and responses that present the assistant as a replacement for relationships with people.
OpenAI also added age-related boundaries covering sexual material, eating disorders, body-image risks, age-restricted goods, dangerous activities, and graphic violence. When a teen may need support, the model is meant to encourage connection with trusted people.
“We reinforced this training with system-level protections and have added new evaluations for how our models perform for users under 18, and are continuing to improve model responses in this area,” the company added.
For jailbreaks, OpenAI tested GPT-5.6 models against adversarial prompts designed to bypass refusal training and produce harmful assistance. The evaluation used challenging multiturn jailbreaks derived from internal red-teaming and excluded additional production safeguards, such as classifiers. OpenAI says GPT-5.6 Sol performed comparably to recent predecessors and was similar to GPT-5.5 Thinking.
For prompt injection, OpenAI tested resistance to attacks that embed adversarial instructions in tool output in an attempt to override system, developer, or user instructions. The tests also included stronger attacks targeting search and function-calling. GPT-5.6 Sol scored 1.000 on the connector evaluation, matching GPT-5.5 and up from 0.998 for GPT-5.4 Thinking. On the separate search and function-calling evaluation, Sol scored 0.910.
