GPT-5.6 Sol vs Luna: Tiers, Slider, 68% Claim

OpenAI swapped the model behind paid ChatGPT conversations on August 6, 2026, moving them onto an updated GPT-5.6 Sol that the company says makes 68% fewer factual errors on its high-stakes evaluations. Free and Go accounts moved onto GPT-5.6 Luna as their default in the same update.

Paid accounts also picked up a reasoning-effort slider and a Think button. The slider and the button put the depth of the model's thinking under the user's hand, one message at a time.

All of it comes from OpenAI's August 6 update announcement, posted with video demos the same day. Both changes land in the ChatGPT app.

Two controls, one dial

OpenAI's framing of the slider is that you choose how much reasoning goes into an answer. Read backwards, that tells you the depth was being chosen for you, message by message, somewhere inside the product. A model can answer immediately, or it can generate intermediate steps before the first visible token appears, and the quality difference on a hard question is large. Until this update, one account could get a sharp answer on Monday and a shallow one on Tuesday with nothing on screen explaining why.

The slider exposes the dial developers already have on the API, where reasoning effort is a request parameter you set yourself. Push it up and the model spends more internal reasoning before it commits to an answer. Push it down and you get speed. The Think button is the single-shot version of the same lever: force the reasoning path on this one message, leave your default where it is.

The control costs something at both ends. High effort means longer waits and heavier consumption of whatever message allowance your plan carries, because reasoning tokens are real tokens even when the interface hides most of them. Low effort on a question that needed the work gets you a fast wrong answer instead of a slow right one. The slider relocates that judgement to the user, and most users will leave it wherever it ships.

Luna's half of the change is an economics move. Luna is the cheap tier of the current GPT-5.6 generation, so making it the free default puts unpaid accounts on current-generation weights. Free tiers run whatever the cheapest current model happens to be. That floor just moved up.

68% fewer factual errors

The figure is OpenAI's, measured by OpenAI, on an evaluation set the company describes as high-stakes. It is a relative reduction, so it compares the updated Sol against the Sol that was answering paid chats before it, on the same tasks.

Relative reductions are the most persuasive shape a factuality number can take and the least legible one. A relative cut of that size tells you how far the error rate moved and nothing about where it started, and where it started decides whether you would notice the difference across a week of ordinary use. If the previous Sol was already right on nearly everything in that set, the improvement is real and invisible. If it was wrong often enough to matter, this is the most useful thing in the whole announcement.

The honest take: OpenAI's 68% tells you the direction of travel and almost nothing about the size of the effect on your own work. The company published the magnitude of the improvement and kept the composition of the evaluation set to itself, so the tasks that produced the number cannot be re-run anywhere outside OpenAI. Direction still counts for something. OpenAI chose to lead a consumer announcement with factuality instead of a coding benchmark, which is a fair read of what actually annoys people about chatbots.

Independent evaluators publish their own cross-model comparisons on their own schedule, and those are the numbers worth waiting for before you conclude the upgrade changed your experience.

August 6 split the tiers

Access splits along the plan lines OpenAI already sells. Plus and Pro, the two paid consumer plans, now get the updated Sol together with both reasoning controls. Free and Go default to Luna and get the Think button. Current plan prices sit on OpenAI's pricing page.

Model and control assignments in the table come from OpenAI's August 6 announcement.

Which one you want follows from the shape of your questions. If your chats are lookups, drafting, summarising and the occasional rewrite, Luna on the free plan now runs the same generation the paid tiers do, and the gap will be hard to feel on that kind of work. If you do multi-step work where a late step depends on an early step being correct, the slider is the thing you are paying for, because reasoning depth is what carries a chain of dependent steps.

What Plus buys you is control over reasoning depth.

There is a second reason to care about the Think button reaching free accounts. Reasoning depth is the single largest quality lever a chatbot exposes to a non-technical user, and handing it to people who have never seen a model parameter is a bigger onboarding change than swapping which model sits behind the box. It also creates a way to get a worse answer than before by turning the dial the wrong way.

Two model IDs on the API

Sol and Luna are also two model IDs on the API, and the comparison there runs on price. Sol is the expensive tier of the family and Luna the cheap one, and we broke down the launch-week sheet in our pricing piece. The gap between the top and bottom of a single model family is the whole argument for routing by difficulty: classify the request before you spend on it, push the easy ones down, keep the flagship for work that fails without it.

The ChatGPT slider is the consumer edition of that routing decision, with the classifier replaced by a person. Developers get a model ID and an invoice they can trace. ChatGPT users get a control they can move and no per-message accounting behind it.

The error rate that OpenAI's 68% reduction was measured against sits in OpenAI's own evaluation logs.