Codex's usage wall is being rebuilt as a cheaper tier, shipped binaries suggest

The restore path carries more information than the fallback itself. According to RuntimeWire's reading of the client, the code holds on to whichever model you had selected, runs you on gpt-reserve, then puts your original selection back when the Reserve window closes ^[3]. A plain downgrade would not need to remember anything. Bookkeeping like that exists so the interruption reads as temporary, and a temporary interruption is the condition under which an offer to buy more premium capacity actually converts.

Compare the two failure modes. A hard limit ends the session and the developer goes to lunch. A metered fallback keeps them in the editor on a weaker model with a visible counter, and the client reportedly ships the string for exactly that counter: "Luna Reserve, {remaining}% remaining" ^[5]. RuntimeWire frames the intent as replacing the wall with a cheaper route that keeps developers active while turning premium capacity into an immediate upsell ^[15]. The mechanism in the binary is consistent with that framing, which is not the same as proving it.

The numbers in the sightings are the part worth putting in a spreadsheet. In the Plus screenshot RuntimeWire reviewed, the account had burned 49 percent of its weekly advanced allowance while the reserve pool sat untouched at 100 percent on a separate reset schedule ^[16][9]. Two allowances, two clocks. A second image from the Codex and Work Analytics view showed ordinary usage at 27 percent against gpt-reserve at 100 percent ^[10]. Once a second pool exists, the published weekly limit stops describing the moment work stops, and starts describing the moment quality drops.

The evidentiary limits are real and RuntimeWire states them. No server interaction or feature-flag bypass was attempted, and live reproduction of the interface still requires server-side eligibility and rate-limit state ^[12]. The gate itself is server-controlled, traced from the flag through eligibility, activation, model switching and restoration ^[11]. OpenAI had not responded to a request for comment by publication ^[13], and the tier does not appear in current official documentation ^[14]. So the client proves that separate metering, Plus and Pro eligibility checks and lifecycle analytics were built and shipped ^[2][1]. It does not disclose how large the reserve pool is, or what it costs.

The work is at least checkable, which is more than the screenshot circuit it started on: a named Store build, a named embedded version and production build number, an archive index of 8,461 entries, plus a published SHA-256 and reproduction steps ^[6][7]. The figure nobody has yet is the ratio of reserve allowance to primary allowance. Until that appears in documentation rather than in a renderer bundle, a Plus seat's effective throughput is whatever a server-side flag says it is this week.