OpenAI Codex deprecation: the date, the reason, and what replaced it

OpenAI announced the Codex deprecation on 20 March 2023 and switched the models off on 23 March 2023. Three days. OpenAI's stated reason, from a customer email that week reported on Hacker News: “Given the advancements of our newest GPT-3.5 models for coding tasks.”

Four model IDs went with it, and that notice window was shorter than any minimum OpenAI publishes today.

Then the name came back in 2025 on a completely different product, which is why this question still gets asked in 2026.

By OpenAI's own account, Codex was a beta that never graduated

Codex launched on 10 August 2021 in private beta. It turned English into code through the completions API, and it never left the status it started in.

OpenAI's own description, in the March 2023 email that ended it: “Codex was initially introduced as a free limited beta in 2021.” That email adds that it held that status to the end. Teams put production traffic on it anyway, which is the part that mattered when the shutdown landed.

Go looking for the launch announcement now and OpenAI tells you it is stale. That 2021 post is still live at its original URL, carrying the line “This launch post is outdated.” That is the whole obituary: a dateline, an outdated notice, and no pointer to what happened next.

We read that dateline out of the raw page markup rather than the rendered page, because a plain fetch of openai.com from anything that looks like a bot returns HTTP 403. Worth knowing if you are building a lifecycle monitor against OpenAI's marketing surfaces instead of its docs. Docs answer. Marketing pages fight you.

Three days is the whole notice window

OpenAI's deprecations page carries the record, under a heading that hands you the announcement date directly: 2023-03-20: Codex models. Its table hands you the shutdown date, and that date is identical on every row.

Table 1 Download CSV

Model ID	Announced	Shut down	Replacement the page lists today
code-davinci-002	2023-03-20	2023-03-23	gpt-4o
code-davinci-001	2023-03-20	2023-03-23	gpt-4o
code-cushman-002	2023-03-20	2023-03-23	gpt-4o
code-cushman-001	2023-03-20	2023-03-23	gpt-4o

That fourth column is a trap. GPT-4o did not exist in March 2023. OpenAI maintains the replacement column as current routing advice, not as a historical record of what customers were told, and the page says as much about itself: it “lists all API deprecations, along with recommended replacements.” Present tense. If you quote that cell as the 2023 migration path, you have published an anachronism.

There is a second trap on the same page, and it also involves code-davinci-002. That ID appears in two separate deprecation entries with two different shutdown dates: 23 March 2023 under the Codex heading, and 4 January 2024 under a later notice dated 6 July 2023, where the listed replacement is gpt-3.5-turbo-instruct. Both entries are real. Both are dated. Nothing on the page explains how they relate, and we are not going to guess, because inventing a story about what happened to that model ID in between is exactly the kind of plausible sentence that gets copied into forty other posts.

Codex shut down on 23 March 2023. January 2024 belongs to a different, later wave.

One more thing about that page, checked on 26 August 2026: there is no prose under the 2023-03-20: Codex models heading at all. Heading, then table. Whatever OpenAI said about why, it does not say it there any more.

The reason OpenAI gave was its own newer models

So the “why” has to come from somewhere else, and the honest version of where is: an email, relayed by a third party.

On 21 March 2023, a Hacker News post titled “OpenAI to discontinue support for the Codex API” reproduced the customer email in full. One sentence carries the reason: “Given the advancements of our newest GPT-3.5 models for coding tasks.” The email then tells customers to move to GPT-3.5-Turbo.

Call it what it is: a contemporaneous secondary source, labelled rather than dressed up. We could not find that wording on any OpenAI-owned page today. Pulling the March 2023 version of the deprecations page out of the Wayback Machine does not work either. Its closest snapshot to the event resolves to a client-side-rendered app shell, so the archived copy holds the JavaScript that would have fetched the text and none of the text. Client-side rendering is a bad thing to be archived as.

For anyone who had Codex in production that week, the practical shape was this: not deprecation-for-cause, not a security retirement. A lab decided that a newer general model was good enough at the specialist's job. It is the most common reason a model dies, and the one that gives you the least warning, because nothing is broken and no incident forces a timeline.

Three days is not a migration window. Three days is a notification that the migration already happened to you.

The replacement has itself been replaced

Follow the chain from March 2023 forward and the destination keeps moving.

Table 2 Download CSV

Date	What happened to the line
2023-03-23	Four Codex models shut down; customers pointed at GPT-3.5-Turbo
2024-01-04	code-davinci-002 retired under a later notice, replacement gpt-3.5-turbo-instruct
2026-02-12	codex-mini-latest removed from the API, announced 2025-11-17
2026-07-23	Five Codex-branded snapshots retired in a 14-model legacy wave

That July 2026 wave is worth reading closely, because it shows what “Codex” costs you as a dependency now. It was announced on 22 April 2026 as a legacy-snapshot retirement covering fourteen model IDs, of which five carry the Codex name: gpt-5-codex, gpt-5.1-codex, gpt-5.1-codex-max, gpt-5.1-codex-mini and gpt-5.2-codex. Four of them route to gpt-5.6-sol. gpt-5.1-codex-mini routes to gpt-5.6-terra. Current rates for those replacements are on our OpenAI API pricing page, and they are not the rates the retired snapshots carried.

Nine other IDs went in the same wave, including search-preview and deep-research snapshots. None of it was a Codex purge. Codex-branded models were simply five of the fourteen things old enough to be swept up, which is a different and slightly worse fact: the name no longer marks a product with its own lifecycle, it marks a variant that ages on the general schedule.

So the 2023 answer to “what replaced Codex” was GPT-3.5-Turbo, and the deprecations page no longer sends anyone there: those four rows now point at gpt-4o. Today's answer is a pair of gpt-5.6 rows that will be replaced in turn. If you are writing a migration doc, write it against the deprecations page rather than against any article, ours included, because the article is a snapshot and the page is the record.

Codex now means something else entirely

Two different products carry this name, four years apart.

OpenAI reused it in 2025 for an agentic coding product built on entirely different models. OpenAI created the openai/codex repository on 13 April 2025 and described it as a “Lightweight coding agent that runs in your terminal.” The public announcement came on 16 April 2025, in the same post that introduced o3 and o4-mini, which called it “Codex CLI, a lightweight coding agent you can run from your terminal.”

One month later, the name expanded again. On 16 May 2025 OpenAI launched a research preview of “a cloud-based software engineering agent,” also called Codex, and stated that “Codex is powered by codex-1, a version of OpenAI o3 optimized for software engineering.” It reached ChatGPT Plus users on 3 June 2025.

Table 3 Download CSV

	Codex, 2021–2023	Codex, 2025–
What it is	Completions models in the API	A terminal CLI and a cloud coding agent
Model IDs	code-davinci-*, code-cushman-*	codex-1, then *-codex snapshots
Announced	10 August 2021	16 April 2025 (CLI), 16 May 2025 (agent)
Status	Shut down 23 March 2023	Shipping, with snapshots retiring on the normal schedule
Who searches for it	Engineers auditing an old integration, or writers checking a date	Engineers evaluating a coding agent this week

Those two audiences type nearly the same query and want opposite pages. If you landed here from “openai codex deprecation” and you meant the CLI, nothing above applies to you: the CLI is alive, and the things that get retired are the model snapshots behind it, on the schedule in the table further up.

Name reuse is not an accident and not a scandal. It is a brand recycled after the thing it named was switched off, which is normal, and which makes four-year-old shutdown facts permanently hard to search for. Readers pay for that, not OpenAI.

A notice window is a policy, not a promise

OpenAI publishes minimum notice periods now. Generally available models get “At least 6 months.” Specialized variants get “At least 3 months,” and the page names Codex variants as one of its examples of a specialized variant.

Set those against the event this article is about. Three days is not close to either number. We are not going to claim OpenAI violated a policy in 2023, because we could not verify that this policy, in this wording, existed in March 2023 — and a rule you cannot date is not a rule you can hold someone to. What we can say is that the 2023 Codex shutdown would be conspicuous under the standard OpenAI publishes today.

There is a more recent case where the arithmetic is worth doing. codex-mini-latest was announced for removal on 17 November 2025 and removed on 12 February 2026. Count it: 87 days. OpenAI's current page says specialized variants, Codex among them, get at least 3 months. Both facts are on the record. We have not verified what the policy text said in November 2025, so we are printing the sum, not a verdict.

Run the same sum on the next wave and it comes out the other way. From the 22 April 2026 announcement to the 23 July 2026 shutdown is 92 days, which clears a ninety-day reading with two days to spare. Two Codex-branded retirements, four months apart, sitting on either side of a boundary OpenAI states in months and does not convert into days. Three calendar months from 17 November 2025 falls on 17 February; ninety days falls on 15 February; removal happened on the twelfth. Which of those readings the rule means is the part the page does not settle.

Compare the way two labs state the same commitment:

Table 4 Download CSV

	OpenAI	Anthropic
Stated minimum	At least 6 months (GA); 3 months (specialized variants)	“60 days' notice before model retirement for publicly released models”
Minimum in days, as we record it	180	60
Publishes forward-looking retirement dates	No	Yes
Lifecycle facts we currently track	81	33

Six months is not automatically the better posture. OpenAI states a six-month floor and does not publish a forward-looking retirement calendar, so you learn about a specific date when the entry appears. Anthropic states sixty days and does publish dates ahead of time, which means a model you depend on can carry a visible expiry before it is urgent. One recent example from our lifecycle records: claude-opus-4-1-20250805 was announced on 5 June 2026 and retired on 5 August 2026, replaced by claude-opus-4-8. Sixty-one days, published in advance, and boring — which is the goal.

Our own coverage has a hole here, and it is the relevant one. We publish per-provider deprecation pages only where every record clears a completeness contract: a retirement date, a replacement, platform scope, and a dated source. Anthropic clears it. OpenAI does not, so there is no published OpenAI deprecation page on this site yet. Provider catalogs marking something “deprecated” are not enough to put a row on a page, and we would rather ship the gap than a table with guessed cells.

How not to get caught by the next retirement

Codex makes a good drill because it removes the two excuses teams usually have. Nobody could argue the model was broken, and nobody had time to argue at all.

• Grep for hard-coded model IDs. Every string like code-davinci-002 in a config, a prompt template, or a fine-tuning script is a dated liability. Find them before someone else's calendar does.
• Watch the deprecations page, not the blog. It is the only surface where the shutdown date and the replacement appear together, and it exposes a Markdown version you can diff on a schedule.
• Name an owner for the alert. Deprecation email goes to whichever address owns the account, which is reliably not the engineer holding the pager.
• Assume the specialist dies first. Codex was a coding model killed by a general model that got good at coding. That pattern has repeated since, and specialist variants carry the shorter stated notice window.

Our guide to monitoring OpenAI API changes covers the mechanics: which surfaces carry which change type, and where diffing beats subscribing. Radar does it across 20 providers, and the reason we run it as a collector rather than a newsletter is structural: a new or missing model ID is an observable fact about an API, while a changelog entry is an editorial decision someone makes later.

Declare five models free — alerts at 90/30/7 days → radar.benchlm.ai/start

Six months is not the number to watch. What matters is the distance between the day a retirement is announced and the day the model stops answering, because that distance is the entire migration budget. For Codex it was three days.