You tune a prompt in ChatGPT until it behaves, paste it into your API call, and the answer comes back longer and mushier. Since 6 August that isn't you imagining things. OpenAI shipped a version of GPT-5.6 Sol tuned specifically for everyday ChatGPT conversation, and it lives only there. The wording in the announcement is unusually blunt: this version will only be available in the Chat experience in ChatGPT, and the Sol powering Work and Codex is not changing. So the model you test in the chat window and the model you bill for now differ on formatting and verbosity, which is most of what prompt work actually is. Free users got the louder half of the news. Unlimited text chats, on a smaller model.
The short answer
On 6 August OpenAI retuned GPT-5.6 Sol for everyday ChatGPT conversation and scoped it to the Chat experience only. Work, Codex and developer traffic keep the model they had. Plus and Pro also get a slider that sets reasoning effort per answer, which quietly retires the Instant and Thinking split. Free and Go drop to GPT-5.6 Luna and gain unlimited text chats. The accuracy claims are real numbers from an eval nobody outside OpenAI has seen.
Image: OpenAI, launch artwork from the 6 August 2026 ChatGPT announcement.
Worth staring at for a second. The artwork is the slider, and it shows five stops with the handle on the third, labelled 5.6 High. The post itself never names a single level.
The fork nobody put in the headline
Every outlet led with unlimited free chats. Fair enough, that’s the consumer story. The line we’d underline sits near the bottom of OpenAI’s post: because this version of GPT-5.6 Sol is optimised for everyday chats, it will only be available in the Chat experience in ChatGPT, and the version powering Work and Codex is not changing.
Read that as an engineer and it’s a fork. Two models, one name, one version string, and no way to tell from a transcript which one answered you. The retune is squarely aimed at output shape: more direct responses, tighter formatting, less detail when detail doesn’t help, and a willingness to correct you instead of agreeing. Those are exactly the behaviours people spend hours prompting around. If your workflow is prototype-in-chat then ship-to-API, your bench and your production just drifted apart, and the drift is invisible because nothing in the model name moved.
Honestly, I don’t think this is sneaky. Tuning a consumer chat surface differently from a developer endpoint is a defensible product call, and OpenAI documented it rather than burying it. But naming both things GPT-5.6 Sol is going to cost somebody an afternoon. We already covered what separated Sol, Terra and Luna at launch, and this is the first time one of those names has stopped meaning one thing.
The practical move is small. Test where you ship. If the answer matters, run the prompt against the endpoint you’ll actually call, not the chat window that happens to be open.
The slider is the real interface change
Plus and Pro no longer pick between Instant and Thinking. One model serves both, and a slider sets how much thought goes into a given answer, on web, mobile and desktop. OpenAI’s framing: moving up should feel like the model taking extra time, not like switching to a different model with its own tone and style.
That framing is the giveaway. Model-switching in ChatGPT has always leaked tone. You’d ask a quick question, get a chatty answer, ask a hard one, and suddenly the voice went formal and bullet-heavy. Collapsing that into an effort dial is the right fix, and it’s one that ChatGPT Work already had before this shipped.
What OpenAI doesn’t publish is how many positions the slider has or what they’re called. The artwork shows five stops and a label reading 5.6 High, which is a design asset rather than documentation, so we’re not going to treat five as a spec. Nothing in the post confirms it.
Unlimited, with a smaller ceiling
Free and Go users get the headline. GPT-5.6 Luna becomes their default that week, and the week after they get unlimited text chats plus a Think button that gives Luna more time on hard questions. Limits stay on file uploads, images and other tools, and the whole thing is subject to abuse guardrails, so unlimited means unlimited typing.
Two things are true at once here, and the coverage mostly picked one. Luna is genuinely better than what free users had: it replaces GPT-5.5 Instant, and OpenAI’s eval puts factual errors 62% lower. That’s an upgrade in quality and a removal of the message cap, on the same day. Good day to be a free user.
The other truth is that Luna is the smallest model in the GPT-5.6 family, and the free tier’s ceiling now sits there unless you press Think. Some readers will file this as free users losing access to frontier reasoning, and that reading holds. I’d put it less dramatically: OpenAI traded depth for volume at the bottom of the stack, which is the obvious move when you’re serving a billion weekly users and inference is the cost line. If you were leaning on the free tier for anything with real stakes, it’s worth checking what you’re actually getting now.
The accuracy numbers deserve one more caveat. 62% and 68% are reductions against GPT-5.5 Instant, a fast non-reasoning model. That’s a low bar by design, since it’s the thing free users were on. It is not a claim that Sol beats a frontier reasoning model on facts, and the announcement doesn’t make that claim either.
What to do with this
If you build on OpenAI, the only action item is the boring one: stop treating the chat window as a staging environment. It was already a rough proxy. Now it’s a different tuning of the same model, and the difference is concentrated in exactly the behaviours you’d notice last, formatting and length.
If you use ChatGPT as a tool rather than a dependency, the slider is a genuine improvement and the merged Instant and Thinking behaviour is overdue. Move it up for anything you’d have hit Thinking for. Leave it low for the rest.
And if you’ve been recommending the free tier to people, recheck that recommendation. The message cap is gone, which is a big deal for casual use. The model behind it is smaller than it was two days ago in family terms, even though it’s more accurate than what it replaced. Both halves matter, and they don’t cancel out neatly. For a wider view of where the family sits on cost, we compared GPT-5.6 against GPT-5.5 when it launched.
Sources
Primary source: OpenAI, Improving GPT-5.6 Sol in ChatGPT and expanding access to GPT-5.6 Luna for free users, 6 August 2026, which carries the chat-only scoping, the accuracy figures and the rollout dates quoted above. Independent coverage confirming the same details: TechCrunch on the free tier limits and what stays capped, 9to5Mac on the Work and Codex carve-out, and The Decoder for the GPT-5.5 Instant baseline and a more critical read of the free tier change.
Frequently asked questions
What changed in GPT-5.6 Sol on 6 August 2026?
OpenAI retuned GPT-5.6 Sol for everyday ChatGPT conversations. It gives more direct answers, uses tighter formatting, adapts detail to the question, and pushes back rather than agreeing when agreeing would not help. Plus and Pro users also got a slider that sets how much thought goes into each answer, on web, mobile and desktop. The retuned model went live for those tiers the same day.
Is the ChatGPT version of GPT-5.6 Sol the same model as the API one?
No, and OpenAI says so directly. The announcement states this version is optimised for everyday chats and will only be available in the Chat experience in ChatGPT, and that the GPT-5.6 Sol powering ChatGPT Work and Codex is not changing as part of the release. The post does not describe any change for developers calling the model through the API. If you prototype prompts in the chat window and ship them elsewhere, you are no longer testing on the model you deploy.
What do free ChatGPT users get now?
GPT-5.6 Luna becomes the default model for Free and Go users, replacing GPT-5.5 Instant. Starting the following week those tiers also get unlimited text chats plus a Think button that gives Luna more time on harder questions, subject to abuse guardrails. Limits still apply to file uploads, images and other tools, so unlimited covers text conversation only.
How much more accurate is it really?
OpenAI reports that in an internal evaluation of financial, medical and legal prompts requiring factual detail, responses containing at least one factual error were about 62% less common with GPT-5.6 Luna and 68% less common with GPT-5.6 Sol than with GPT-5.5 Instant. Those are the company's own numbers, with no published prompt set, sample size or grading method, and no independent reproduction. The baseline is also a fast non-reasoning model, so the comparison flatters the result.
What happened to the Instant and Thinking modes?
For Plus and Pro they are merged. One model now serves both quick answers and higher effort, and the slider moves between them. OpenAI frames the goal as making higher effort feel like the model taking more time rather than like switching to a different model with its own tone. The slider was already present in ChatGPT Work before this release.
