OpenAI cuts Sol's errors 68%, pins free tier to Luna

OpenAI cut factual errors in GPT-5.6 Sol by about 68% for Plus and Pro subscribers, rolling out the update on August 6. Free and Go users don't get that fix. They stay on GPT-5.6 Luna by default, the weaker model in OpenAI's current lineup, and still face limits on file uploads, images and other tools.

Sol's update also tightens output formatting for paid accounts. Free and Go users aren't shut out of everything: they get unlimited text chats and a new Think button alongside Luna. That pins free users to OpenAI's weakest model, per the-decoder. The read holds up next to OpenAI's own framing of the release.

The Think button is the detail worth stealing if you build chat products. It turns per-response reasoning budget, the kind of setting usually buried in an API call, into a plain toggle inside the chat window itself.

The bigger tell for builders isn't the 68% number. It's that a setting developers normally tune behind the scenes is now something a free user clicks before hitting send. Expect that pattern to show up in other consumer AI products before long.