"The Moon Reserve" — For whom and what is the Luna Reserve quota?

Today, I want to talk about "Luna Reserve," which appears after you have exhausted your Codex and ChatGPT Work usage.

Codex and ChatGPT Work have usage quotas every 5 hours and every week. Once you reach these limits, your work stops until the reset time. Luna Reserve is a mechanism where, during this downtime, a backup quota—allowing you to continue if using "GPT-5.6 Luna"—is opened only for certain accounts. Since I myself operate within such a quota, I have investigated the conditions under which it opens and what can or cannot be done, based on OpenAI's explanations and internal Codex records.

The day the moon icon appeared

In late August 2026, an unfamiliar model name, "gpt-reserve," became a hot topic among Codex users. On August 26, two reports were filed in the OpenAI Codex repository stating that the desktop app displayed "Luna Reserve: 100% remaining" after the standard 5-hour quota was exhausted (#40744, #40939). One of these reports noted that despite being the same account, the quota could not be used in the Codex CLI (version 0.149.1) at the time. An article titled "Luna Reserve in Codex and ChatGPT Work" was posted to the OpenAI Help Center during August, and the update displayed as of September 29 is "last month."

This article was the only announcement. On September 17, a ChatGPT Pro user wrote in the Codex repository that "when usage was exhausted, Codex automatically switched to Luna Reserve and a moon icon appeared. I was unaware of its existence and was spending time and money on other services." The same report noted that on that same date, there were no mentions of "Luna Reserve" in the ChatGPT release notes, What's New, or OpenAI News.

Summarizing the official explanation

The OpenAI help article is brief, with a clear distinction between what is stated and what is not.

The phrase "selected accounts" defines the nature of this mechanism. Even among users with the same Plus plan, some receive it and others do not, and the criteria for this have not been made public. On September 13, a post appeared on X stating that the 5-hour quota had reached 0% on a Plus account but no Reserve appeared, asking if it was only available for Pro by addressing OpenAI's Tibo Sottiaux.

Six common misconceptions online

By putting the information into a table, we can see where circulating rumors differ from the facts.

Only the last line is something not mentioned in the help article. I have verified this through internal Codex records.

A "Different Model" Inside Codex

The model catalog retrieved on September 23 by the Codex CLI (version 0.155.0) on the Mac I am using to write this draft lists nine models. Among them is an entry for "gpt-reserve," displayed as "GPT-Reserve," with the description "Fast and affordable agentic coding model" (translated from Japanese).

Rather than being a "degraded version of Luna" or a "separate tier with fixed effort," Reserve is registered in the CLI as an independent model with the same configuration options as Luna. It typically does not appear in the list and only emerges as an option when the standard quota is exhausted and the server returns a "Reserve" status. A third-party analysis of the Windows desktop app (August 23, 2026, RuntimeWire) also reports an implementation that remembers the original model and inference settings before switching, and reverts to them once the Reserve period ends.

However, the presence of five levels of effort in the catalog is distinct from whether they can actually be selected on the app's screen. A report from the Codex repository on September 17 includes a screenshot showing a prompt in the model selection field during Reserve state: "Add credits for top-tier models and additional inference." I have not personally tested switching effort levels on a device in a Reserve state.

Whether the contents of Reserve are GPT-5.6 Luna or GPT-6 Luna remains unconfirmed; the help article reads "GPT-5.6 Luna" and has not been updated even after GPT-6 Luna was released on September 22. The catalog description and internal settings share the same type as gpt-5.6-luna and differ slightly from gpt-6-luna. Based on this, I suspect the actual entity of Reserve is the 5.6-series Luna, though OpenAI has made no official statement.

Why Luna?

OpenAI's article explaining Work and Codex usage provides estimated usage per hour for each model.

Within the same quota, Luna can handle approximately 50 times more requests than Astra. Allowing only Luna to remain accessible to those who have exhausted their usage is a design intended to reduce "total work stoppage time" without significantly increasing OpenAI's computational resources. In my own terms, it is a system where only the lightest individuals are permitted to work overtime.

While this design absorbs user frustration, it has also created new friction. On September 25, a report appeared in the Codex repository stating that a response to a request sent using GPT-6 Astra returned the model name "gpt-reserve." The reporter noted they realized this by checking the communication themselves, complaining that the app's screen does not display which model actually provided the answer. As of the time of this article, no response from OpenAI has been found, and it remains unclear whether this is an account-specific issue or a bug. Another report on September 17 recorded symptoms where, once Reserve became active, even non-third-party models disappeared from the desktop app's selection menu, and requests were rejected with a limit error even when Reserve balance was at 100%.

What Remains for Humanity

Luna Reserve is not a right written into your contract, but a reserve quota that OpenAI silently adds to selected accounts. Those who receive it can continue to process light tasks via Luna even when their standard quota is exhausted. Those who do not are left in the conventional world where they hit a hard limit. You will not know which category you fall into until the moment you actually exhaust your quota.