OpenAI Shipped a Slider for How Hard the Model Thinks. That's a Pricing Decision.

On August 6, 2026, OpenAI shipped a slider. ChatGPT Plus and Pro users can now choose how much reasoning effort the model spends on each response, with the control available across web, mobile and desktop — low for quick questions, higher for research, planning, coding and decisions with more moving parts. Free users moved to GPT-5.6 Luna as their default model with unlimited text chats, and get a discrete Think button rather than a continuous slider.

This will be covered as an interface change. It is a cost architecture surfacing as a control, and the difference between the paid slider and the free button is the clearest statement anyone in this industry has made about where inference margin actually comes from.

Reasoning made inference cost variable

For most of the deployed history of large language models, the cost of answering a question was roughly a function of how long the answer was. Input tokens in, output tokens out, price per token, done. It was predictable enough that vendors could sell flat-rate consumer subscriptions without much anxiety.

Reasoning models broke that. A model that deliberates before answering — generating intermediate reasoning that the user never sees — can consume wildly different amounts of compute on two questions that produce identically short answers. 'What is the capital of France' and 'Should I restructure this debt' both return a paragraph. One of them can cost a hundred times more to produce than the other.

That variance has to be managed somewhere. Until now it was managed invisibly, behind a model picker: the vendor chose which model handled which request, absorbed the variance, and presented the user with names rather than costs. Enterprise products solved the same problem more explicitly with model routing — a cheap model handles the easy traffic, an expensive one handles the hard traffic, and a classifier decides which is which.

The gap between a slider and a button

The tiering is the tell. Plus and Pro get a continuous control. Free gets a binary one.

If reasoning effort were purely a user-experience preference, there would be no reason to withhold granularity from free users — a slider is not more expensive to render than a button. It is withheld because a continuous control lets a user park it at maximum and hold it there, and the cost of that behaviour is unbounded in a way a per-request button is not. A binary Think toggle caps the worst case and makes each expensive request a deliberate act.

Pair that with the other half of the free-tier change — unlimited text chats on GPT-5.6 Luna — and the shape of the offer becomes legible. Unlimited volume on a cheaper default model, with expensive reasoning gated behind an explicit per-request action. That is a coherent unit-economics position, and it is worth noting that not all coverage read it as generosity: at least one outlet framed the same change as restricting free users to OpenAI's weakest model. Both readings describe the same product decision. Which one you find persuasive depends on what free users were getting before.

Reading the 68% claim carefully

OpenAI's headline reliability figure for this release is specific, and it is worth quoting precisely rather than paraphrasing, because the paraphrases circulating are not equivalent to the claim.

“In an internal OpenAI evaluation covering financial, medical and legal prompts requiring detailed facts, responses containing at least one factual error were 68% less common with GPT-5.6 Sol than with GPT-5.5 Instant.”

Four qualifiers are doing real work in that sentence. It is an internal evaluation, meaning OpenAI selected the prompts, ran the grading, and reported the result; there is no external audit and no public benchmark to reproduce it against. The metric is the share of responses containing at least one factual error — not an error rate per claim, and not a measure of how severe the errors were. The prompt set is domain-specific: finance, medicine and law, chosen because they are fact-dense, which is a defensible design choice and also not representative of typical usage. And the comparison is against GPT-5.5 Instant, OpenAI's own previous fast model.

That last point is the one most worth sitting with. A comparison against your own prior model establishes a direction of travel. It does not establish a level. If GPT-5.5 Instant produced errors in a large fraction of responses on a hard fact-dense prompt set — which, being the fast tier, it plausibly did — then a 68% reduction still leaves a meaningful residual error rate. The figure tells you the new model is better than the old one on this set. It tells you nothing about how it compares to a competitor, or whether it is now reliable enough to use unsupervised on a legal or medical question.

None of this makes the claim dishonest. Vendor-run evaluations against a prior version are standard practice and genuinely informative. The error is in the translation: '68% fewer responses containing factual errors than our own previous fast model, on our own prompt set' becomes '68% more accurate' in the retelling, and those are not the same statement. The second one is not supported by anything OpenAI published.

The same caution applies to the release's softer claims — that the model calibrates detail to the question, uses less unnecessary formatting, and pushes back rather than agreeing when agreement would be unhelpful. These are OpenAI's characterisations of its own product's behaviour. They may well be accurate. They are not measured, and there is no way for a reader to check them except by using the thing.

How to actually use the slider

Practical guidance, given what reasoning effort does and does not buy:

• Raise it for problems with multiple interacting constraints — planning, debugging, comparing options where the trade-offs affect each other. This is where deliberation genuinely changes the answer.
• Raise it when the failure mode is a subtle logical error rather than a missing fact. More thinking catches contradictions; it does not conjure information the model never had.
• Leave it low for retrieval, formatting, summarising, rewriting, and anything where you already know what the answer should look like. Extra reasoning here buys latency and nothing else.
• Do not raise it to fix hallucination about facts the model cannot know. A model reasoning harder about a figure it has never seen will produce a more confident wrong figure, not a right one.
• Treat maximum effort as a cost, not a default. If you cannot articulate what the extra deliberation is for, it is probably just waiting.

The pattern, not the product

The reason this is worth more than a product note is that it is not one company's design choice. Effort dials, thinking budgets and reasoning toggles have been appearing across consumer AI products through 2026, and they are all responses to the same underlying economics: reasoning made the cost of a request unpredictable, and unpredictable cost is incompatible with flat-rate pricing at scale.

There are only a few ways out of that. Absorb the variance and accept margin risk. Route silently and accept that users cannot tell which model answered them. Meter usage explicitly and accept that consumers dislike it. Or surface the dial and let the user spend their own allocation — which is the option that preserves flat-rate pricing, and the one the industry appears to be converging on.

The consequence for users is subtle but real. When the vendor chose the model, the vendor owned the quality of the answer. When you choose the effort level, part of that ownership moves to you. A slider is a feature, and it is also a transfer of responsibility.