ChatGPT Reasoning Slider Puts Thinking Effort in Your Hands

OpenAI Hands ChatGPT Users a Dial for How Hard the Model Thinks

OpenAI shipped an updated version of GPT-5.6 Sol into ChatGPT on August 6, 2026, and paired it with a control the interface has never offered before: a slider that lets you decide how much reasoning effort the model spends before it answers. It is a small piece of user interface with a surprisingly large effect on how the product feels, because for the first time the tradeoff between speed and depth is a choice you make rather than a decision the router makes on your behalf.

• The updated GPT-5.6 Sol reached ChatGPT Plus and Pro subscribers on August 6, 2026, across web, mobile, and desktop
• A reasoning effort slider offers five settings, from Instant through Medium, High, and Extra High up to Pro
• In an internal OpenAI evaluation on financial, medical, and legal prompts, responses containing at least one factual error were 68% less common than with GPT-5.5 Instant
• Free accounts move to GPT-5.6 Luna with unlimited text chats and a new Think button for harder questions

What Does the ChatGPT Reasoning Slider Actually Do?

Every modern reasoning model spends a variable amount of computation between reading your prompt and producing a reply. That budget is what determines whether the model plans, checks its own work, and explores alternatives, or simply answers from its first instinct. Until now, ChatGPT decided that budget internally. The slider exposes it.

Drag it toward Instant and you get the behaviour that suits a quick factual lookup or a casual back-and-forth: fast, tight, low latency. Drag it toward Pro and the model is allowed to spend substantially longer working through the problem before it commits to an answer, which is the setting that pays off on planning, research synthesis, long-form writing, and multi-file coding work. The five stops — Instant, Medium, High, Extra High, Pro — give enough granularity to match effort to task without turning the choice into a configuration exercise.

The design point worth noticing is that this is a per-conversation control, not a buried setting. OpenAI is betting that users have a better sense of how much a given question is worth than any automatic router does, and on the evidence of how people actually use these tools, that is a reasonable bet.

Why Effort Dials Are Becoming a Standard Control

This is part of a broader convergence across the frontier labs. Reasoning models introduced a genuinely new axis of product design: the same weights can behave like a fast assistant or a slow analyst depending purely on how much thinking you buy. Hiding that axis makes the product feel inconsistent, because two identical-looking prompts can produce very different quality. Exposing it makes the behaviour legible.

It also has a practical cost dimension. Reasoning tokens are the expensive part of an inference bill, and giving users a visible dial is the cleanest way to let them spend deliberately rather than uniformly. We saw a similar economics-driven move when OpenAI cut Luna pricing after AI-written GPU kernels improved its serving efficiency, a story we covered in our OpenAI Luna price cut analysis. Effort dials attack the same problem from the demand side instead of the supply side.

How Much More Reliable Is GPT-5.6 Sol?

The headline number from OpenAI's own testing is a 68% reduction in responses containing at least one factual error, measured against GPT-5.5 Instant on prompts drawn from financial, medical, and legal domains. Those three categories are a deliberate choice: they are detail-dense, they reward precision over fluency, and they are exactly where a confident-sounding wrong answer does the most damage.

Two caveats are worth stating plainly. This is an internal evaluation rather than an independent benchmark, and the comparison point is GPT-5.5 Instant, which was the fast tier rather than the deepest one. Neither of those undercuts the result, but they do frame it: the improvement is largest precisely where quick answers used to be weakest. Alongside the accuracy work, OpenAI tuned the conversational model to give more direct responses with tighter formatting and less padding when extra detail does not help — a change that is harder to benchmark but easy to feel.

What Free Users Get

The free tier moves to GPT-5.6 Luna with unlimited text conversations, replacing the previous message caps, plus a Think button that pushes a single question into deeper reasoning when it needs it. Unlimited text chat on a free tier is a meaningful expansion of access, and it follows the same direction as OpenAI's decision to open its research program to academic users, which we wrote about in our piece on free ChatGPT access for researchers.

What This Means for Everyday Use

The practical advice is simple. Leave the slider low for the majority of what you do, because most questions genuinely do not need extended reasoning and the latency cost is real. Push it up deliberately when the answer will be acted on: a financial model, a medical summary you intend to discuss with a professional, a refactor across several files, a research brief someone else will rely on. Treating the slider as a per-question decision rather than a global preference is where the value is.

For a broader look at how reasoning models are evolving, see our ongoing artificial intelligence coverage.

Sources: OpenAI — August 6, 2026; 9to5Mac — August 6, 2026; Digital Watch Observatory — August 2026.