ChatGPT Thinking Slider: What Changed and Who Gets It
OpenAI moved the reasoning-effort dial that API developers have had for years into the ChatGPT interface, and reshuffled what free and paid tiers get. The tradeoff is more volume for less capability.
OpenAI announced on 6 August 2026 that ChatGPT is getting a thinking slider for paying users, a Think button for free users, and an updated GPT-5.6 Sol. Plus and Pro subscribers can now dial how much effort the model spends on an answer. Free and Go users move to GPT-5.6 Luna as their default, gain unlimited text chats, and get a Think button the following week.
Almost every write-up covered this as a model update. It is more accurately a control update, and the two halves of it point in opposite directions.
What each tier gets
Tier | Default model | Effort control | Volume |
|---|---|---|---|
Free and Go | GPT-5.6 Luna | Think button, runs Luna at higher effort | Unlimited text chats |
Plus and Pro | GPT-5.6 Sol, updated | Thinking slider, reported as five settings | Existing limits |
The important detail in that table is the third column of the first row. TechCrunch reported that the Think button runs GPT-5.6 Luna, not Sol. It extends how long the smaller model thinks; it does not promote you to the larger one.
Separate caps stay in place for files, images, voice and image generation, so unlimited applies to text conversations only. GPT-5.6 Sol inside Codex and the work products is unchanged.
The thinking slider is an old idea arriving in a new place
Developers calling these models through an API have had a reasoning-effort parameter for a while. You set it low for classification and extraction, high for planning and hard code. The tradeoff is money and latency against quality, and you tune it per call rather than per account.
The ChatGPT thinking slider is that same parameter with the cost hidden and the labels rewritten for people who do not think in tokens. The Decoder reported five settings, spanning everyday questions at one end and planning, research and coding at the other.
That is a real usability win, and it also quietly teaches a hundred million people something the difference between reasoning models and standard ones usually has to be explained to them: more thinking is not free, and most questions do not need it.
The accuracy claim, and how much weight to put on it
OpenAI says responses containing at least one factual error were 68 percent less common for GPT-5.6 Sol and 62 percent less common for Luna, compared with GPT-5.5 Instant. The figure comes from an internal evaluation on finance, medicine and law prompts.
Read that carefully. It is a relative reduction against a specific older model, on a prompt set OpenAI chose, scored by OpenAI, with no published methodology and no independent replication. The Decoder flagged exactly this. A 68 percent relative drop is compatible with a very large or a very small absolute change depending on where the baseline sat, and the baseline is not given. Treat it as a claim of direction, not a measurement you can plan against.
That is also the right standard to hold any new model release to: before trusting a headline accuracy number, it is worth reading the model's system card to see what was actually evaluated, rather than taking a summary blog post's word for it.
The behavioural notes in the announcement are easier to verify yourself: Sol is meant to cut unnecessary detail and excess formatting, and to push back where simply agreeing would be unhelpful. That last one is a direct swipe at sycophancy, which until now has mostly been something you had to prompt around.
The part worth arguing about
Free users got more conversations and a smaller default model in the same announcement. Luna is the smallest model in the 5.6 family. It is stronger than the GPT-5.5 Instant it replaces, so nobody is worse off than last week, but the free tier no longer reaches a frontier reasoning model at all, which it previously could.
That is a defensible business decision at roughly a billion weekly users, and it is also a structural change in what "I asked ChatGPT" means. Two people running the same query now get answers from models a tier apart, and neither is told which. If you are shipping a product and using ChatGPT to sanity-check outputs, note which account you were on.
For builders the practical follow-on is unchanged by any of this: pricing and capability tiers keep moving, so keep a written process for evaluating a model change rather than reacting to each announcement, and keep reading the primary release notes rather than the summaries. The price cut on GPT-5.6 earlier this cycle is the other half of the same strategy.
How did this land?
About the author
Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.
