Superpower Daily
OpenAI Loses Data Center Chief While Pursuing a Reported $600 B...Alibaba Raises HK$80B for Full-Stack AI as Investors Question t...LLMScholarBench Tests 22 Models and Finds an Accuracy-Represent...Meta Adds Voice-Controlled AI and Navigator Home Screen to Ques...Harvard Uses 7 AI Faculty Avatars in a $699 Founder Bootcamp, W...DeepSeek V4 Pro Lands on Fireworks With $2.50 CyberGym Solves,...OpenAI Loses Data Center Chief While Pursuing a Reported $600 B...Alibaba Raises HK$80B for Full-Stack AI as Investors Question t...LLMScholarBench Tests 22 Models and Finds an Accuracy-Represent...Meta Adds Voice-Controlled AI and Navigator Home Screen to Ques...Harvard Uses 7 AI Faculty Avatars in a $699 Founder Bootcamp, W...DeepSeek V4 Pro Lands on Fireworks With $2.50 CyberGym Solves,...
Productspublished

OpenAI Codex Client Shows Luna Reserve Fallback After Advanced-Model Limits

The client code describes a temporary, separately metered path for eligible Plus and Pro accounts, but a server-side feature gate means the code alone cannot show who receives it.

By 2 min read
OpenAI Codex Client Shows Luna Reserve Fallback After Advanced-Model Limits
Sign in to save and react

Listen to this story

The audio brief

About 1:26
0:001:26
Read transcript
OpenAI’s Codex Windows client is built to keep a session going after a user hits an advanced-model limit. The code describes a temporary fallback called Luna Reserve, which can switch an eligible Plus or Pro account to a separate model, gpt-reserve, with its own allowance and rate limit. When that reserve period ends, Codex is designed to restore the model the user had selected before the handoff. That makes this more than a new balance display. The client preserves the user’s preference, substitutes the reserve model only after the advanced-model allowance is exhausted, and then returns to the original choice. Its interface includes “Luna Reserve — {remaining}% remaining,” while the production Windows package contains a reserve_enabled feature gate, Plus and Pro checks, and lifecycle analytics tied to the feature. Reported images also show GPT-Reserve in Codex’s model selector. The important limitation is that the client code does not prove broad access. The gate is controlled on OpenAI’s servers, so eligibility, rate-limit state, rollout size, and the actual allowance terms can all be determined remotely. In other words, the code shows a designed workflow and its intended audience, not which accounts can use it today. OpenAI had not commented at publication time. The key question now is whether Luna Reserve is activated for any particular account, and how much separately metered capacity those users receive.

Story brief

3 key points

OpenAI appears to be preparing a fallback capacity path in its Codex Windows client for Plus and Pro users who exhaust advanced-model limits. The internally named Luna Reserve uses gpt-reserve, tracks a separate allowance and rate limit, and returns users to their prior model afterward. Evidence comes from Windows package 26.818.5229.0 and Codex build 26.818.41509, but the feature remains server-gated. The key...

  1. 01

    The build includes reserve_enabled, Plus/Pro checks, lifecycle analytics, and a “Luna Reserve · {remaining}% remaining” balance display.

  2. 02

    Reserve usage is metered independently, separating fallback capacity from the exhausted advanced-model allowance.

  3. 03

    Codex is designed to restore the user’s previous model when the reserve period ends.

OpenAI’s Codex Windows client outlines a softer usage cutoff. When an eligible subscriber reaches an advanced-model limit, the app is designed to switch to an internally named reserve model with its own allowance. It then restores the user’s previous model selection when the reserve period ends. The feature, called Luna Reserve, remains behind a server-side gate.

The reserve handoff

The sequence is more than a new meter in the interface. The client is designed to preserve the model a user had selected, temporarily substitute gpt-reserve after the advanced-model allowance is exhausted, and return the original selection after Reserve ends. That structure would let Codex continue a session under a different model path without requiring the user to reset their preference.

A client-side trail

The production Windows build includes the internal model identifier gpt-reserve, a reserve_enabled feature gate, Plus and Pro account checks, and lifecycle analytics for Luna Reserve. It also contains interface text for a balance display: “Luna Reserve · {remaining}% remaining.” Together, those elements indicate a distinct product path rather than an unused model name.

The findings were drawn from Windows package 26.818.5229.0 and its embedded Codex app, version 26.818.41509, build 6962. Reported images also show GPT-Reserve in Codex’s model selector, alongside a separately metered allowance. Together, the interface and code indicate that Reserve was built to be visible to at least some enabled accounts.

Access is still server-controlled

The client does not establish that every Plus or Pro subscriber can use Luna Reserve. Its feature gate is server-controlled, and live access depends on server-side eligibility and rate-limit state. The subscription checks show the intended audience for the feature, but not its rollout size, its allowance terms, or whether access is broadly enabled.

A different kind of limit

The notable product choice is the separation between premium-model capacity and the fallback path. A user who hits an advanced-model limit would not simply encounter the same limit under a different label: the reserve model has independently tracked usage. Restoring the original model afterward keeps that fallback from becoming the user’s permanent selection.

OpenAI had not responded to a request for comment at publication time. For now, the Windows client establishes a designed reserve workflow and its eligibility controls, while the decisive question is whether OpenAI turns that workflow on for an individual account.

Related coverage

Continue reading