Free ChatGPT Is Unlimited Now. The Catch Is Called Luna.

TL;DR

On August 6, OpenAI reshuffled ChatGPT's model lineup without shipping a single new model. Per its announcement, GPT-5.6 Luna becomes the default for Free and Go accounts this week, and the cap on standard text messages disappears next week alongside a Think button that escalates a hard question to deeper reasoning. Plus and Pro get a retuned GPT-5.6 Sol: tighter formatting, less padding, and an explicit instruction to correct you rather than nod along. OpenAI's internal evals claim responses containing at least one factual error fell about 62% for Luna and 68% for Sol versus GPT-5.5 Instant. Your API keys, Codex, and ChatGPT Work are untouched.

What actually changed, by surface

This is a product change, not a release. Same weights, same model names, different defaults and different knobs depending on what you pay.

OpenAI is explicit that the Sol retune lives in ChatGPT only. The Decoder reports the update "only applies to ChatGPT" and that Sol "remains unchanged in ChatGPT Work and Codex." If you have Sol wired into an agent loop, your evals do not need rerunning.

Unlimited, on the smaller model

The two halves of this announcement are not independent. Free accounts get uncapped text conversations, and free accounts also get moved onto Luna, the cheapest tier in the family. Those are the same sentence read from two directions.

ChatGPT crossed roughly a billion weekly users this month, per Stocktwits' readout of the announcement. Removing a message cap at that scale is not a product decision, it is a capacity decision, and it only pencils out if the marginal conversation is cheap enough to be a rounding error.

Why unlimited became affordable a week ago

Seven days before this, OpenAI cut Luna's API price 80%, from $1.00 / $6.00 to $0.20 in and $1.20 out per million tokens, which the pricing page still lists. We covered that cut when it landed. The serving economics that justified the API discount are the same economics that make an uncapped free tier survivable.

It is also a shot at the free tiers of Claude and Gemini, both of which still meter conversations. "Unlimited" is a very hard word to compete with on a pricing page, even when the asterisk is a model name.

Where the 62% and 68% come from

OpenAI's stated numbers are a reduction in the share of responses containing at least one factual error, measured against GPT-5.5 Instant on finance, medicine, and law prompts. The Decoder's quote of the announcement: responses with at least one factual error "fell about 62 percent for GPT-5.6 Luna and 68 percent for GPT-5.6 Sol."

Read that metric carefully, because it is a binary per response, not a count of wrong claims. It grades essays pass or fail on whether they contain a single typo: you learn how many came back clean, and nothing at all about how badly the failures failed. A model that halves its clean-response rate while doubling the severity of the errors it does make would look terrific on this chart.

These are internal evals with no published methodology, no eval set, and no external replication. The Decoder flags the same thing. Treat them as a direction of travel, not a measurement.

Sol was told to stop nodding along

The more interesting half is behavioral. OpenAI describes the retuned Sol as producing more focused answers, tighter formatting, less unnecessary detail, and corrections in cases where simply agreeing would not be useful. That last clause is a sycophancy fix stated out loud in release copy, which is new.

Alongside it, Plus and Pro get a slider for how much thought each response gets. Reasoning effort has been an API parameter for a while; this is the first time OpenAI has handed the dial to consumers in the chat product. The free tier gets the same idea in one-bit form: the Think button.

Reporting differs on how many positions the slider has, so we are not going to invent a number. What matters is the shape: OpenAI has spent two years training people that the model picker exists, and is now betting that most of them never open it.

One thing the coverage does not agree on

Whether Free users lose access to Sol entirely, or simply get Luna as the new default with Sol still reachable, is reported both ways. The Decoder says free users "lose access to OpenAI's most powerful reasoning altogether." Developers Digest says they do not, and that Luna is simply the new default.

We could not reach OpenAI's post directly to settle it, so we are not asserting either version. For the overwhelming majority of free users the practical outcome is identical, because the default is the model they will use. If you are on a free account and you care, check your model picker before you plan around it.

If you ship on the API, nothing moved

No new endpoint, no new model ID, no price change, no deprecation. Sol in the API is the same Sol as yesterday. The signals worth filing away are the second-order ones.

• The cheapest tier in a frontier family is now good enough to carry a billion-user consumer default. That is a statement about small-model quality, not about generosity.
• Reasoning effort has graduated from an API parameter to a consumer UI control, which means users will start arriving with opinions about how hard your product thinks.
• Metered free tiers just got harder to defend. If your product competes for casual usage against a chat box with no message cap, the comparison is now blunt.

Key Takeaways

• Free and Go accounts default to GPT-5.6 Luna this week, with the standard text message cap removed and a Think button arriving next week.
• Plus and Pro get a retuned GPT-5.6 Sol plus a reasoning-effort slider, explicitly scoped to the ChatGPT chat surface.
• OpenAI claims responses with at least one factual error fell about 62% for Luna and 68% for Sol against GPT-5.5 Instant, on internal finance, medicine, and law evals with no external replication.
• The API, Codex, and ChatGPT Work are unchanged. No model IDs, prices, or behavior moved for anything you deploy.
• Luna's 80% API price cut on July 30 is the enabling event. Uncapped free chat at roughly a billion weekly users only works if the marginal conversation is nearly free to serve.
• Reporting conflicts on whether Free users keep any Sol access. Verify in the product rather than trusting a summary, this one included.

Sources: OpenAI, "Improving GPT-5.6 Sol in ChatGPT", The Decoder, The Next Web, Developers Digest, Stocktwits, OpenAI API pricing