The undocumented feature gives some Plus and Pro users a separate allowance after advanced-model access is constrained. RuntimeWire traced its model switching and metering logic inside OpenAI’s Windows client.
By Ryan Merket · Published · Updated
RUNTIMEWIRE INVESTIGATION — Scoop
Original reporting by RuntimeWire, based on reverse engineering.
Why it matters #
Reserve would let OpenAI replace Codex's hard usage wall with a cheaper fallback, keeping developers active while turning premium capacity into an immediate upsell.
Reporting record #
Finding
Codex contains a feature-gated “Luna Reserve” allowance that appears to switch eligible Plus and Pro users to a Luna-class reserve model after advanced-model limits are reached.
How we verified
Methods: reverse engineering.
Static analysis of the production-distributed Codex app.asar identified the model identifier gpt-reserve, feature gate reserve_enabled, luna_reserve banner logic, dedicated lifecycle analytics, eligibility checks for Plus/Pro accounts, and user-facing text describing a separate “Luna Reserve” allowance. The code preserves the user’s original model selection and restores it when Reserve ends.
Calculated the archive’s SHA-256 hash, parsed its ASAR index, and extracted the Electron main-process and renderer bundles. Searched the bundles for feature definitions, model identifiers, rate-limit handling, eligibility conditions, analytics events, and user-facing strings. Traced Luna Reserve from its server-controlled feature gate through account eligibility, rate-limit activation, model switching, UI presentation, and restoration of the original model. Compared findings against current official OpenAI documentation. No server interaction or feature-flag bypass was attempted.
Tested versions: Windows package 26.818.5229.0, embedded Codex app 26.818.41509, build 6962 (prod).
Reproduction
RuntimeWire independently reproduced the core finding.
Obtain the specified Microsoft Store build of Codex and hash app/resources/app.asar. Parse the ASAR and inspect the renderer bundle for gpt-reserve, reserve_enabled, luna_reserve, and CodexLunaReserveLifecycle. Confirm the surrounding code includes Plus/Pro eligibility checks, a separate gpt-reserve rate limit, model switching, and the user-facing string “Luna Reserve · {remaining}% remaining.” Live UI reproduction additionally requires server-side eligibility and rate-limit state.
File hashes
C5D839BC9B122B7EF2A2F0F45186B3E5895923DE5B6CEF5253C936FE670C0479
Company response
RuntimeWire requested comment; the company had not responded by publication time.
OpenAI has begun rolling out an undocumented fallback tier for Codex that can keep subscribers working after their regular advanced-model allowance runs low.
The feature is called Luna Reserve. It carries the internal model identifier gpt-reserve
, maintains its own usage allowance and is designed to temporarily replace a user's selected model while preserving the settings needed to switch them back later.
Multiple images posted to X and reviewed by RuntimeWire show the feature surfacing in user accounts. One shows GPT-Reserve in Codex's model selector, directly beneath GPT-5.6 Luna and carrying the same visible "Fast and affordable agentic coding..." description. Others show gpt-reserve
as a separately metered allowance alongside an account's ordinary weekly limit.
In one image, a Plus account had 51 percent of its standard weekly allowance remaining while gpt-reserve
displayed 100 percent on a different reset schedule. The user said the unfamiliar allowance had "appeared today" in their quota display. Another image from OpenAI's Codex and Work Analytics interface showed ordinary usage at 27 percent and gpt-reserve
at 100 percent.
Those sightings establish that Reserve is no longer merely dormant code. RuntimeWire's examination of OpenAI's production Windows client reveals what the screenshots cannot: who the feature is designed for, how it is activated, how it changes models and how OpenAI keeps paid routes back to its more capable models within reach.
A second tank, not a renamed gauge
RuntimeWire examined the Codex package distributed through the Microsoft Store as version 26.818.5229.0
. The application's Electron archive contained 8,461 entries and reported an embedded app version of 26.818.41509
, production build 6962.
Inside the client, Reserve is treated as an additional rate limit - not a new name for the subscriber's existing allowance.
The application searches account-usage data for an entry whose limit_name
is exactly gpt-reserve
. It tracks that allowance separately, calculates its remaining capacity and renders dedicated Reserve progress indicators. User screenshots showing two balances with different reset schedules match that implementation.
The client also recognizes a backend banner type called luna_reserve
. That response can tell the app when to expose Reserve and which account actions to display alongside it.
Eligibility is tightly controlled. The code requires a ChatGPT-authenticated individual account, excludes workspace accounts and checks for plan identifiers corresponding to Plus, Prolite and Pro. It also requires a recent local Codex runtime - version 0.142.2
or newer - and two layers of server-controlled feature gating, including a flag named reserve_enabled
.
That architecture gives OpenAI several levers over the rollout. It can decide which accounts receive Reserve, when the client exposes it and which commercial options accompany it without shipping another desktop update.
Codex is built to switch models - and switch back
The most consequential code is not the progress bar. It is the model-transition logic.
When the Reserve state is entered, Codex can record the user's original model and reasoning configuration before moving the session to gpt-reserve
. When Reserve ends, the client restores the previous configuration.
That makes Reserve more than a usage-dashboard experiment. It is a continuity system built to preserve an active workflow while moving it onto a separately metered model.
The client includes dedicated lifecycle events for entering and exiting Reserve, errors, banner impressions, calls to action, header interactions and modal activity. Reserve can appear in several parts of the interface, including limit banners, the model picker and a persistent sidebar meter reading:
Luna Reserve - {remaining}% remaining
Another message tells users when "Advanced models reset," drawing an explicit distinction between Reserve and the advanced-model allowance it supplements.
The production code does not establish that every eligible user will be switched automatically. Activation still depends on server-provided account and rate-limit state. But it shows that OpenAI built the client to perform the switch, preserve the original settings and reverse the change when the fallback period ends.
The Luna connection is strong - but not conclusive
Reserve's relationship with GPT-5.6 Luna appears throughout the implementation and the surfaced interface.
A client-side model test groups gpt-reserve
with gpt-5.6-luna
. The feature is named Luna Reserve internally, uses a moon symbol and, in a user-posted model selector, carries the same visible description as GPT-5.6 Luna.
That is powerful evidence that Reserve receives Luna-class treatment. It is not proof that the two identifiers invoke identical model weights or infrastructure. The final routing decision occurs on OpenAI's servers, beyond what can be established from the desktop client.
The economics fit the design. OpenAI describes GPT-5.6 Luna as optimized for "cost-sensitive, high-volume workloads" and says it roughly corresponds to the nano tier in earlier GPT-5 families. Its listed API price is $0.20 per million input tokens and $1.20 per million output tokens - one-tenth of GPT-5.6 Terra's current rates. OpenAI's GPT-5.6 Luna documentation
OpenAI's broader model guidance positions Luna as the efficient member of a three-tier family: Sol for frontier capability, Terra for balance and Luna for high-volume work. OpenAI's GPT-5.6 model guidance
A lower-cost fallback would let OpenAI preserve session continuity without continuing to serve every depleted subscriber on its most expensive capacity. It would also give users a direct experience of the difference between Reserve and the advanced model they had selected.
That is the obvious economic interpretation. Until OpenAI identifies the backend behind gpt-reserve
, it remains an inference.
A soft landing with an upgrade button
Traditional usage limits create a hard stop. Reserve creates a softer boundary.
Instead of simply ending a user's work when advanced-model capacity is constrained, OpenAI can move the session into a separately metered fallback. The user keeps working, but the interface continues to advertise routes back to "advanced models and added reasoning."
Those routes are selected by the backend and vary by account. The Reserve interface can render a primary action, a secondary action and, where supplied, a localized purchase price. The underlying account-action system supports options involving additional usage, resets and subscription changes.
The result is both a continuity feature and a conversion surface.
A user who would otherwise leave at the limit can remain inside Codex. At the same time, the product places the difference between fallback and premium capacity directly in that user's workflow - and keeps restored access one account action away.
Reserve may therefore solve two problems for OpenAI at once: retaining work that would have stopped at the quota boundary and creating a more immediate reason to pay for renewed advanced-model access.
Live, but still unexplained
As of August 23rd, OpenAI's public desktop documentation does not contain the terms "Luna Reserve" or gpt-reserve
. OpenAI's current desktop-app documentation
RuntimeWire requested comment from OpenAI and will update this story when the company responds.
The company has not publicly explained:
- How much Reserve usage eligible accounts receive.
- Whether the allowance is standardized or individually assigned.
- Whether activation is automatic for every eligible user.
- Why standard and Reserve allowances can have different reset dates.
- How Reserve performance compares with GPT-5.6 Luna and the user's original model.
- Whether
gpt-reserve
is technically identical to Luna or a distinct serving configuration. - How large the current rollout is.
- Whether Reserve will remain included, become paid or disappear after testing.
The evidence nevertheless supports a clear conclusion.
OpenAI has built and begun deploying an undocumented fallback system for Codex. It is separately metered, restricted by plan and account type, controlled from the server, connected to rate-limit state and capable of replacing a subscriber's selected model while remembering how to restore it.
Reserve is not a secret frontier model hiding inside Codex. It is potentially more consequential: a new layer in the economics of AI subscriptions, placed precisely where a user's included access ends and the decision to pay again begins.
Methodology
RuntimeWire performed a read-only static examination of the production app.asar
distributed with the Windows Codex package. The archive was indexed and relevant Electron main-process and renderer bundles were extracted and searched for feature definitions, account eligibility, model identifiers, rate-limit handling, interface strings and lifecycle events.
The package measured 286,869,415 bytes and contained 8,461 ASAR entries.
Windows package: 26.818.5229.0
Embedded app: 26.818.41509
Build: 6962 (prod)
SHA-256: C5D839BC9B122B7EF2A2F0F45186B3E5895923DE5B6CEF5253C936FE670C0479
Static examination cannot reveal OpenAI's server-side model weights, every rollout rule or the complete population receiving the feature. RuntimeWire compared the implementation with images posted by users showing GPT-Reserve
in the model selector and gpt-reserve
as a live, separately metered allowance.