From 6 August, paid users get a retuned model with a slider for how deep it thinks. Free users switch to GPT-5.6 Luna by default this week, and get unlimited text conversations from next week. Files and images stay capped.
- The free tier gets the stronger model by default and unlimited text chat. The limits on files, images and tools stay.
- By OpenAI's internal evaluation, answers with at least one factual error are about 62-68% rarer than the previous model. There's no named public benchmark.
- The change is only in the chat screen. The versions doing the work in Codex aren't touched.
From 6 August, free ChatGPT switches to GPT-5.6 Luna by default, with unlimited text chat. OpenAI does two things at once: retunes the model for paying users and lowers the threshold for free ones.
You'll see the 62-68% number retold as "68% more accurate". It isn't that. What's measured is how much rarer it is for an answer to have at least one error, against a specific previous model, on the company's own internal test. There's no named public exam, no outside checker. The direction is probably right. The magnitude is their word.
One excuse I hear regularly falls away - 'I don't have a subscription'. Have you hit a limit after four questions? Now you can actually get real work done, without paying. Running a training, or showing something to a client? The entry bar just dropped. The difference for us is small - nothing moves in Codex. Everyday work stays what it was.