OpenAI has announced significant updates to its ChatGPT service, including the removal of text chat rate limits for free users and the introduction of enhanced safeguards for younger individuals. The company is rolling out new models, GPT-5.6 Sol and GPT-5.6 Luna, which aim to improve accuracy and provide more focused responses.
Free and Go tier users are being migrated to GPT-5.6 Luna as their default model, with text chat rate limits for these users being eliminated starting next week. While text conversations will become unlimited, other functionalities such as file uploads and image processing will retain their existing limits. Free users will also gain a new "Think" button for more in-depth reasoning on complex queries.
For Plus and Pro subscribers, the updated GPT-5.6 Sol model is now available. This model integrates both quick replies and longer reasoning into a single system, replacing the previous Instant and Thinking modes. OpenAI states that this change is designed to deliver more precise answers, adapt detail levels to the specific question, and avoid unnecessary formatting. A new slider allows these users to adjust the reasoning effort the model applies to a response. The version of Sol powering Work and Codex remains unchanged.
OpenAI reports a notable reduction in factual errors with the new models. Internal tests on prompts related to finance, medicine, and legal topics showed that answers containing at least one factual error were approximately 62 percent less common with Luna and 68 percent less common with Sol, compared to the outgoing GPT-5.5 Instant model.
The company has also expanded its safeguards for users under 18 with GPT-5.6 Sol. The model has been specifically trained to avoid romantic roleplay, age-restricted challenges, and presenting itself as a substitute for human relationships. New age-related boundaries have been implemented to address issues such as sexual material, eating disorders, body image risks, age-restricted goods, dangerous activities, and graphic violence. When a teen might require support, the model is designed to encourage connection with trusted individuals. OpenAI has reinforced this training with system-level protections and new evaluations for model performance with younger users.
In terms of security, OpenAI tested the GPT-5.6 models against adversarial prompts designed to bypass refusal training, known as jailbreaks. The evaluation utilized challenging multi-turn jailbreaks derived from internal red-teaming exercises and excluded additional production safeguards like classifiers. GPT-5.6 Sol performed comparably to its recent predecessors and was similar to GPT-5.5 Thinking in these tests.
For prompt injection attacks, which embed adversarial instructions in tool output to override system, developer, or user instructions, GPT-5.6 Sol demonstrated strong resistance. In the connector evaluation, Sol scored 1.000, matching GPT-5.5 and improving upon GPT-5.4 Thinking's score of 0.998. On a separate evaluation for search and function-calling, Sol achieved a score of 0.910.
