Three days was the whole migration budget

OpenAI announced the Codex deprecation on 20 March 2023 and switched the four models off on 23 March 2023. The stated reason survives only in a customer email quoted on Hacker News that week: newer GPT-3.5 models had gotten good enough at coding. Nothing was broken, so nothing forced a timeline. This week we published the full record — the dates, the model IDs, the traps on OpenAI’s own deprecations page, and why the name now means a different product entirely. The distance between the announcement and the shutdown is the entire migration budget. For Codex it was three days.

What the record actually says

Three days, four model IDscode-davinci-002, code-davinci-001, code-cushman-002, and code-cushman-001 were announced dead on March 20 and gone on March 23. Shorter than any minimum OpenAI publishes today.The replacement column is a trapOpenAI’s deprecations page now lists gpt-4o against the Codex rows. GPT-4o did not exist in March 2023 — the column is current routing advice, not a historical record.Codex the model is not Codex the agentThe 2021 completions API stopped answering in 2023. The 2025 CLI and cloud agent reuse the name on different models, which is why two audiences typing the same query want opposite pages.The next wave already happenedFive Codex-branded snapshots retired on 23 July 2026 inside a 14-model legacy sweep. Four route to gpt-5.6-sol, one to gpt-5.6-terra — at different prices than the retired IDs carried.

The limit

The reason for the 2023 shutdown comes from a contemporaneous secondary source — a customer email reproduced on Hacker News — because no OpenAI-owned page carries that wording today, and the Wayback copy of the deprecations page is an unrendered app shell. We also publish per-provider deprecation pages only where every row has a retirement date, a replacement, platform scope, and a dated source; OpenAI’s record does not clear that contract yet, so there is no BenchLM OpenAI deprecation page to link you. The Radar counts below are live API values captured on the send date.

Radar since the last issue

Declare five models free

One free account, no card. Declare up to five models and Radar emails you when a published retirement date crosses 90, 30, and 7 days out. The morning Brief arrives at 09:00 ET only on qualifying mornings.

A public lifecycle API, no key

Summary counts, a calendar.ics your calendar app can subscribe to, RSS and JSON feeds, and a full-dataset calendar.json for CI. Published lifecycle facts only, with per-route dates and a dated source behind each row.

The route is part of the date

The same model can carry a different retirement date on Bedrock or Azure than first-party. Radar records dates per serving route and currently tracks 18 models whose routes disagree.

Pro routes the alert, not just sends it

Per-channel notification profiles across email, Slack, Discord, and signed webhooks, plus remote MCP tools that let an agent read the retirement runway and manage the declared stack without opening the app.

Analysis worth opening

OpenAI Codex deprecation: the date, the reason, and what replaced itThe full receipt trail: the three-day window, the double retirement of code-davinci-002, the anachronistic replacement column, and the notice-window arithmetic on both 2026 waves.How to monitor OpenAI API changesThe smallest monitoring stack for announcements, deprecations, pricing, incidents, callable IDs, and detailed changelog entries — and where diffing beats subscribing.

This issue in numbers

303
Models with tracked lifecycle facts: 81
Retiring within 90 days: 151
Already past a published date: 18
Models whose routes disagree on the date

New on BenchLM

Start with the free Radar stackDeclare up to five models and get the 90/30/7 runway alerts with no card. Radar Pro is $19.99/month for the complete event stream across every supported provider; current pricing and trial eligibility appear before checkout.

Browse BenchLM

Tools & calculators

Tools