ChatGPT Goes GPT-5.6: Free Tier Gets Unlimited Luna

OpenAI announced on August 6, 2026 that GPT-5.6 Luna will become the default model for Free and Go users in ChatGPT this week, and that those users will get unlimited text chats plus a new Think button the week after, subject to abuse guardrails. It is the largest single expansion of what a non-paying ChatGPT account can do since the GPT-5.6 family launched, and it is easy to describe wrongly.

The wrong description is already circulating: “ChatGPT is unlimited now.” It is not, on three counts. Unlimited applies to text chats only, with file uploads, images and other tools still capped. It applies to Free and Go, not to Plus or Pro. And at the time of writing it has not shipped — OpenAI frames it in the future tense, for the week of August 10.

This guide takes the announcement apart in the order it actually happens: the three separate effective dates folded into one post, what changed for each ChatGPT surface, what OpenAI’s two accuracy figures do and do not claim, the API price trajectory that makes unlimited plausible as an economic decision, and what any of it should change for teams whose audiences now default to a free-tier assistant.

OpenAI published the change on August 6, 2026 under a title that names both halves of it: improving GPT-5.6 Sol in ChatGPT, and expanding access to GPT-5.6 Luna for free users. The same text appears verbatim in the ChatGPT release notes on OpenAI’s help centre, which is the canonical dated changelog and the version worth checking if you want to know what state your own account is in.

The paid half is a quality and control change. Plus and Pro users get an updated GPT-5.6 Sol in the Chat surface, which OpenAI describes as more reliable with facts and as giving more focused answers, along with a slider that lets them choose how much thought ChatGPT puts into a response. Both landed on announcement day across web, mobile and desktop.

The free half is an access change, and it is the one that moves the market. Free and Go accounts move to GPT-5.6 Luna as the default model in regular chats, and are then scheduled to receive unlimited text chats and a Think button that gives Luna more time to work through an answer. Between the two halves sits a carve-out that is easy to miss: because this version of Sol is tuned for everyday chats, it is only available in the Chat experience. OpenAI states that the version of GPT-5.6 Sol powering Work and Codex is not changing as part of this release.

Most coverage of this change compresses it into a single present tense. The announcement does not support that. It contains three distinct effective dates, and the difference between them is the difference between a fact and a forecast.

The Sol update and the effort slider are described as available now, on August 6. Luna as the Free and Go default is described as happening this week. Unlimited text chats and the Think button are described as starting next week, which puts them in the week of August 10. At the time of writing, on August 7, only the first of those three has an announcement date behind it that has already passed.

That matters more than it sounds. If you are writing documentation, briefing a client, or building a comparison table, a sentence like “Free ChatGPT users now have unlimited chats” is not accurate on August 7. The precise version is that OpenAI has said they will, starting next week, and has attached an abuse-guardrail caveat to the promise. Keep the vendor’s own tense.

The scope question has a plain answer in the source, which is rare enough to be worth quoting rather than summarising. OpenAI’s release-notes entry reads: “GPT-5.6 Luna will become the default model for Free and Go users this week. Starting next week, they’ll also have unlimited text chats and access to a new Think button for harder questions (subject to abuse guardrails). Limits will still apply for file uploads, images and other tools.”

Three constraints are stated there and each of them is load-bearing. The unlimited applies to text chats, so the moment a conversation involves an upload, an image generation, or another tool, a limit is back in play. It is subject to abuse guardrails, which is the standard vendor reservation permitting rate limiting on patterns that look automated or abusive. And it applies to the Free and Go tiers named in the sentence, which does not extend upward to Plus or Pro.

The last of those is the misreading most likely to end up in a client deck. Pro is the most expensive consumer plan, so the intuition is that whatever Free gets, Pro already has. Here the direction of travel is the opposite: Free gets a cap lifted on one interaction type, while Pro gets a better model and a new control with no cap change described at all. Anyone drafting a plan comparison this month should read our full comparison of the ChatGPT tiers rather than reasoning from price alone.

Every piece of coverage we read describes the after state. Almost none reconstructs the before state, which is where the change actually becomes legible. Before August 6, Free and Go users did have access to a GPT-5.6-family model — Terra — but only inside ChatGPT Work and Codex. In regular chats they remained on the older default. Axios documented that split in July, and it is the cleanest baseline available for judging the size of this week’s move.

The table below is our own reconstruction across six surfaces. The before column for Free and Go comes from that July reporting; the after column and every date come from OpenAI’s August 6 post and the matching release-notes entry. Where the announcement is silent, the cell says so rather than guessing.

Read down the after column and the shape of the release is obvious. OpenAI moved the floor, not the ceiling. The biggest capability jump in the announcement belongs to the accounts that pay nothing or eight dollars a month, while the paid tiers get a refinement and a control surface. That is a deliberate distribution decision, and it is the opposite of how most software vendors sequence a model upgrade.

It is also the second free-tier expansion from OpenAI in three months. In June, personalization drawing on past chats and connected accounts reached Free and Go — a different feature with the same demographic target, which we covered in our guide to ChatGPT personalization arriving on the free tier. Two data points is not a strategy, but the direction is consistent: close the gap at the bottom faster than anyone expects.

The most quoted number from this announcement is 62%. It is real, but it travels badly, because it is one of a pair and because it is narrower than the shorthand suggests. OpenAI’s exact wording is: “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.”

Unpack that and four qualifiers appear. It is an internal evaluation, with no independent audit published alongside it. The prompt set is financial, medical and legal prompts requiring factual detail, which is a deliberately hard, fact-dense slice rather than general usage. The measured quantity is responses containing at least one factual error, which is a response-level pass or fail rather than an error count. And the baseline is GPT-5.5 Instant specifically — the model being replaced, not the current best available option.

The two figures also belong to two different models on two different tiers. The 62% is Luna, the model Free and Go are moving to. The 68% is the updated Sol, which Plus and Pro get. Fusing them into a single “about 65% fewer errors” claim, or attaching the 68% to the free tier, misstates who gets what. The chart below indexes both against the shared baseline so the relationship stays visible without the two numbers collapsing into each other.

One caution on reading that chart. The bars are an index of a relative reduction, not an absolute error rate — OpenAI did not publish what share of GPT-5.5 Instant responses contained an error in the first place. A 62% reduction from a small number and a 62% reduction from a large number look identical here and mean very different things in practice. For anything with a compliance or liability surface, the operative fact remains that a factual-error rate above zero is still a factual-error rate, on a model now serving as the default for the largest user cohort OpenAI has.

The paid half of the announcement is smaller in headline terms and more interesting in product terms. Plus and Pro get an updated GPT-5.6 Sol in the Chat surface, described by OpenAI as more reliable with facts and as giving more focused answers, and framed as bringing Instant and Thinking together into one consistent experience rather than two modes a user has to pick between.

Alongside it comes a slider that lets Plus and Pro users choose how much thought ChatGPT puts into a response, live from August 6 on web, mobile and desktop. This is a genuinely notable interface move: reasoning effort has been an API parameter for developers for some time, and this pulls the same idea into the consumer chat window where the person paying for the tokens can see the trade-off they are making. The free tier gets a two-state version of the same idea in the Think button, which OpenAI describes as giving Luna more time to work through the answer.

The pattern is spreading across vendors fast enough to deserve its own treatment, which is why we wrote a separate piece on what these effort dials actually buy you. The short version: an effort control changes how long the model deliberates, not what it knows. It is a latency-and-cost dial with a quality correlate, not a model swap, and treating it as the latter leads to disappointment on questions that need retrieval rather than reflection.

A companion system card published with the release documents new under-18 safety measures, including avoiding romantic roleplay and age-restricted challenges, not presenting the model as a substitute for real-world relationships, age-appropriate boundaries around sexual content, eating-disorder and body-image risk, age-restricted goods, dangerous activities and graphic violence, and encouragement to connect with trusted people. That documentation exists partly because an unlimited free tier changes the composition of who is using the product and for how long, which is a reasonable thing for a vendor to plan around before it happens.

OpenAI did not explain why unlimited became possible in August rather than July. The following is our own reading of the timeline, not a vendor-stated reason, and it should be treated as analysis: in the week before the unlimited announcement, the API list price of a Luna token fell by four fifths.

On July 30, 2026, OpenAI cut GPT-5.6 Luna’s API list price by 80% and Terra’s by 20%. The table below tracks the developer-surface standard list price per million tokens across that change. Every percentage in the final column is recomputed from the two prices on the same row rather than restated from the announcement, so the arithmetic is checkable. One caveat that matters: these are API rates for developers, not what any ChatGPT subscriber pays. They illustrate the cost structure behind a consumer decision; they are not the consumer decision.

The recomputed ratio is the part worth holding onto. Before the cut, Sol cost five times what Luna cost on both input and output — $5.00 against $1.00, and $30.00 against $6.00. After the cut, the same comparison is twenty-five times: $5.00 against $0.20, and $30.00 against $1.20. In one week, the list-price gap between OpenAI’s flagship and its high-volume model widened by a factor of five without the flagship moving at all.

That is the economic shape in which an unlimited free tier stops being reckless. If the cost of a free-tier text turn moves anything like the list price does, the cost of removing the counter falls with it, and the counter itself starts costing more in friction and support than it saves in compute. We track the wider set of August rate moves and their expiry dates in our AI API pricing cuts and promos tracker, which is the place to check whether a given number is a permanent list change or a promotional rate with a reversion attached.

Two honest caveats. First, none of this is stated by OpenAI as a reason — the causal link between the July 30 cut and the August 6 unlimited decision is our inference from two dated public facts, and a reader is entitled to reject it. Second, API list prices are not internal serving costs; a vendor’s list price reflects strategy as much as unit economics. What the trajectory shows for certain is that OpenAI decided Luna should be cheap, and then decided it should be everywhere. Those two decisions are one week apart.

OpenAI’s own framing for the change is reach. Its post states that every week, one billion people turn to ChatGPT for everything from quick questions and web searches to planning, research, advice and complex decisions. That is a vendor-stated figure and it is the scale at which removing a text-chat counter becomes a meaningful infrastructure commitment rather than a feature flag.

The context around it is less comfortable. Mobile-app data from analytics firm Apptopia, reported this week, indicates that ChatGPT’s share of United States daily active users among generative-AI chatbot apps was broadly lower in July 2026 than in November 2025, while Claude and Meta AI gained share and Gemini remained a major app with a declining share of its own. The same data set shows that globally, downloads of the most popular generative-AI chatbot apps recorded their first negative-growth month in July 2026.

Those figures are qualitative in the reporting we could verify — broadly lower, first negative-growth month — with no precise percentage attached, and we are not going to invent one. Treat them as directional. The framing that OpenAI expanded the free tier to defend share against rivals is press analysis rather than a stated motive, and it is worth labelling as such. But the timing is at minimum a coincidence worth noticing: the category posts its first negative-growth month on this data, and within weeks the category leader says it will remove the main reason a casual user would keep a second assistant installed.

For most teams the operational impact of this release is smaller than the headline and larger than zero. Nothing in it changes an API integration, an agent, or a Codex workflow — the carve-out is explicit. What changes is the capability floor of the audience: from the week of August 10, the modal person asking an AI assistant about your category is doing it on a better model, with no conversation counter telling them to stop.

The forward-looking read is about where the competitive boundary moves next. If uncapped text conversation becomes table stakes at the free tier across vendors — and a category leader removing a counter tends to force that — then the differentiators left for paid consumer plans are the things this announcement kept capped: uploads, image generation, tool access, context length, and the degree of control a user has over how hard the model thinks. That is a narrower and more functional set of levers than “more messages,” and it is a harder one to explain on a pricing page. Expect the next year of consumer AI plan design to be fought over tools and context rather than turn counts.

For brands, the practical consequence is that assistant answers stop being a truncated first impression and start being a sustained conversation. A user who could ask three questions before hitting a wall got a summary of your category. A user who can ask thirty gets a considered opinion, formed largely from whatever the model can retrieve about you. Our agentic SEO engagements exist precisely for that problem, and the broader question of where AI fits in an operating model is what our AI transformation work is built around. If you want the model-family background behind Luna, Terra and Sol, start with our coverage of the GPT-5.6 GA launch in July and the earlier three-tier preview guide.