The 20-day figure below is this publication’s own arithmetic from the date of an entry on OpenAI’s deprecations page. The page does not say 20 days, states no notification date, and gives no reason for the length of notice.This publication did not test the API. Nothing below reports a breach of OpenAI’s policy, which carries an exception for safety or compliance concerns.
OpenAI’s own deprecations page shows an entry for gpt-5.4-cyber dated September 11, 2026, with a shutdown date of October 1, 2026. The 20-day interval is computed by this publication. The page does not state it, and it gives no notification date for this entry.
This publication fetched and read OpenAI’s deprecations page directly at developers.openai.com/api/docs/deprecations. One entry reads, verbatim: “the gpt-5.4-cyber model is deprecated and will be removed from the API on October 1, 2026. Migrate to the most capable cyber model available to you before the shutdown date.”
The entry is dated 11 September 2026. The shutdown date listed is 1 October 2026. This publication computed the interval as 20 days. The page does not state 20 days, and it does not give a notification date for this entry. This publication has not tested the API and did not confirm that the model has in fact been removed.
The same page carries a policy section. It states, verbatim: “unless safety or compliance concerns require a faster timeline, we provide the following minimum notice periods before model retirement. Generally available models: at least 6 months. Specialized variants of generally available models: at least 3 months.” The page also notes that preview models may be retired with much shorter notice, such as 2 weeks, and identifies that tier by the word “preview” in the model name. The name gpt-5.4-cyber does not contain the word preview.
The key insight: Twenty days is short of both the three-month and the six-month minimums stated on the page. The page does not classify gpt-5.4-cyber into a tier, does not say the exception was invoked, and does not say it was not. The entry gives no reason for the notice length.
The page’s stated minimums are three months for specialized variants and six for generally available models. Day counts are this publication’s own arithmetic.
The Structural Read
The policy is written with a standard and an exception. The standard is the minimum notice period. The exception is safety or compliance concerns. The page carves out that exception twice.
The first carve-out attaches to the minimums themselves. The second reads, verbatim: “If safety or compliance concerns require us to retire a model sooner, we will provide as much notice as reasonably possible.” That second sentence contains no number.
This is an observation about how the policy is written, not a judgment about this entry. The planning-relevant figure for a developer is not the three or six months. It is whatever the exception permits. The page does not give that figure.
Permission Layer Read
A floor with an uncapped trapdoor
A published minimum notice period creates developer expectations. An exception clause without its own floor resets those expectations to an undefined lower bound. The two exist in the same document. Developers reading for planning purposes encounter both simultaneously.
The page carries 36 dated entries going back to 2022. This publication compared only the 2026 model entries listed above and makes no claim that 20 days is the shortest notice on the page or in OpenAI’s history.
The comparison that the page itself makes possible is narrow but concrete. The entries dated between June and August 2026 gave removal dates roughly six months out: 182, 183, and 184 days computed. The entries dated 1 October 2026 for gpt-5.3-codex, gpt-5.1, and gpt-5.4-nano give 182 days, and the page describes those as six months’ notice. Notably, the page identifies gpt-5.3-codex as a Codex variant, which it places in the specialized-variant tier with the three-month minimum — yet it received six months. The text-to-speech models on the same date received 97 days, which the page describes as at least three months’ notice.
Against that pattern, the gpt-5.4-cyber entry at 20 days sits apart. The page does not explain why. The entry gives no reason.
OpenAI Deprecations Page — Policy Section
“Unless safety or compliance concerns require a faster timeline, we provide the following minimum notice periods before model retirement. Generally available models: at least 6 months. Specialized variants of generally available models: at least 3 months.”
What the Page Leaves Out
The page states that when OpenAI announces a deprecation, it notifies customers who are actively using the model by email. It gives no date on which that email went out for gpt-5.4-cyber.
The recommended replacement is also notable for what it does not name. The instruction is to migrate to “the most capable cyber model available to you.” That is a description, not a model identifier. A developer cannot tell from the page which model that is.
The page notes that in some cases, developers may be able to provision dedicated capacity for continued access after a shutdown date, and that the route is to contact the sales team. It does not say whether that option exists for gpt-5.4-cyber.
Not established, and therefore absent from this article: why the notice was short, which tier the model falls in, who had access to it, how many developers used it, when any email was sent, whether the removal happened on schedule, and whether OpenAI has said anything about this entry elsewhere.
Three Implications
IMPLICATION 1 — THE POLICY READS DIFFERENTLY THAN IT PLANS The stated minimums — three months and six months — are the numbers developers naturally plan around. But the exception clause, which contains no minimum of its own, is the operative constraint when invoked. The page does not say how often or under what circumstances that happens.
IMPLICATION 2 — VAGUE REPLACEMENT GUIDANCE COMPOUNDS THE PROBLEM A migration instruction pointing to “the most capable cyber model available to you” without naming a model identifier leaves developers to resolve the replacement themselves. With 20 days on the clock, that ambiguity has an operational cost that a named successor would not.
IMPLICATION 3 — THE DEDICATED-CAPACITY OPTION IS UNRESOLVED FOR THIS ENTRY The page describes dedicated capacity as a route for continued access past a shutdown date, conditional on contacting the sales team. For gpt-5.4-cyber, the page neither confirms nor excludes that option. Developers who needed to explore it had 20 days in which to find out.
Business Engineer Framework
The Permission Layer
The Permission Layer framework maps how governance decisions — including a provider’s own published policies — shape what developers can build and for how long. This entry is a live case: a published floor, an uncapped exception, and a 20-day window computed from a documentation page. Understanding where that layer sits in the AI stack is how builders avoid building on sand.
OpenAI’s deprecations page says the minimum is three or six months, carves out an exception with no number attached, and carries an entry for gpt-5.4-cyber where this publication computed 20 days between the dated entry and the listed shutdown. The page gives no reason, names no tier for the model, identifies no replacement by name, and gives no date for any notification email. Whether the minimums applied, whether the exception was invoked, and whether the model was actually removed on October 1 are all things the page does not settle — and that this publication did not independently verify.
Source: OpenAI Developers — Model Deprecations Page, fetched and read directly by this publication on 2 October 2026. All day-count intervals are computed by this publication from dates on that page. Nothing in this article is investment advice.
Every quotation and date above comes from OpenAI’s own API deprecations page, read directly by this publication on 2 October 2026. This publication did not test the API and did not confirm that gpt-5.4-cyber has in fact been removed; the page says it will be removed on 1 October 2026.The figure of 20 days is this publication’s own arithmetic: the interval between the entry dated 11 September 2026 and the shutdown date of 1 October 2026.
The page does not say 20 days and states no notification date for this entry. The other day counts are computed in the same way.The page gives no reason for the length of notice and does not say which notice category gpt-5.4-cyber falls in. Its policy carries an exception for safety or compliance concerns, and nothing above says that exception was or was not used.
Nothing above reports a breach of OpenAI’s policy.The comparison covers only the 2026 model entries named above. The page carries 36 dated entries going back to 2022, and nothing above claims that 20 days is the shortest notice on it or in OpenAI’s history.Nothing above says who had access to the model, how many developers used it, or when any email was sent. Nothing above predicts anything and nothing here is investment advice.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
