Summary: On August 6, OpenAI retuned GPT-5.6 Sol for paid ChatGPT and made GPT-5.6 Luna the default for free and Go users with uncapped text chats and a Think button. Internal finance, medicine, and law evaluations versus GPT-5.5 Instant found at least one factual error in 62% fewer Luna answers and 68% fewer Sol answers. The update followed July 30 API price cuts and reached a user base OpenAI put at one billion weekly people.
The largest ChatGPT product change OpenAI has made this year landed on August 6: paying users get a retuned GPT-5.6 Sol, while free users see the default model switch to GPT-5.6 Luna with text chats no longer capped. It arrives just three and a half weeks behind the July 30 API price cuts for the GPT-5.6 family, and it reaches a user base OpenAI puts at one billion people each week. The change lands on both extremes of willingness to pay at once: those paying twenty dollars a month and those paying nothing get experiences that differ, yet both improve.
Instant Replies and Deep Reasoning in a Single Model
Up to now, ChatGPT’s quick replies and its heavier reasoning came across as two separate products: a different tone, a different manner, and a jolt of disorientation whenever a question crossed the line between them. This update merges both experiences into one GPT-5.6 Sol for Plus and Pro users. The way OpenAI frames it: the feeling should be that the model spends more time thinking, not that you have moved to another model with a personality of its own.
The one thing users touch is a slider, present alike on web, mobile, and desktop, that decides how much thought ChatGPT gives to any single reply: keep it low for everyday questions, raise it for planning, research, writing or code, or for decisions that warrant more care. Sol has also been tuned to hold back more; it matches detail to the question at hand, keeps formatting to what the question calls for, and, where agreeing with you would serve no purpose, corrects you instead.
Factual Errors Drop: 62% and 68%
Reliability draws the heaviest emphasis, and it is the only place in the announcement where hard numbers appear. An internal evaluation drew on prompts in finance, medicine, and law that demand factual detail; measured against GPT-5.5 Instant, the previous default, answers carrying at least one factual error turned out to be some 62% rarer with Luna and 68% rarer with Sol. OpenAI credits the model’s better use of the sources it retrieves when answering.
For the ordinary user this comes down to something simple: confident wrong answers grow rarer. For someone who works with ChatGPT as a productivity tool, this counts for more than any benchmark. Finance, medicine, and law — the three areas OpenAI tested — are where a fabricated number or an invented deadline wreaks real damage, and likewise where casual users increasingly look for guidance.
Free and Go Tiers: Luna as the Default, Uncapped Text, a Think Button
As of this week, free users find GPT-5.6 Luna as their default — a small, speed-focused one that only debuted in early July — and text chats come with no cap. The inexpensive Go subscription gets identical treatment. And when a question calls for greater depth, a newly added Think button opens up higher reasoning for the harder problems, so free users still have a path up to the strongest end of the stack.
Axios cut straight to the point: free plans from the frontier labs now give away more intelligence than most people can ever use up. What sets paid tiers apart is narrowing down to speed, headroom, and reach into the strongest reasoning — whether the product works at all is no longer the question.
Why Give So Much Away, and Why Now
The timing reads easily enough. OpenAI had just cut API prices for Luna and Terra on July 30, and Axios traces the pressure back to inexpensive Chinese open-weight models — a force that keeps pushing OpenAI and Anthropic to show their closed models earn their premium. Serving a speed model such as Luna without limits is a cost that stays manageable: it was designed to be cheap to run, and text is the least expensive modality to give away. In return, the case for Plus shifts from “it works” to “it works better” — sharper answers, one consistent voice, and a single slider that scales effort rather than swapping personas.
Two lessons stand out for product teams. First, one model with adjustable thinking effort is becoming the standard interface for consumer AI — the practice of shipping two models with two personalities inside one app is winding down, and users will come to expect effort as a setting rather than a product boundary. Second, a free tier is no longer a warning about a crippled version; it is a customer acquisition funnel. Assume your users can always get “just good enough” intelligence for free, and that whatever you charge for has to beat free in a way they can feel within one session.