The code model thousands had built on was announced dead on a Monday and gone by Thursday.

OpenAI's Codex models were the first widely available systems for turning English into working code, and by early 2023 they sat underneath a large number of developer tools, side projects and published research. They had been free the whole time, in a limited beta with heavy rate limits [2].

On 20 March 2023 OpenAI announced that `code-davinci-002`, `code-davinci-001`, `code-cushman-002` and `code-cushman-001` were being retired. The shutdown date was 23 March 2023 — three days later [1].

Users were told to move to GPT-3.5 Turbo, the model then powering ChatGPT, which OpenAI said would improve programming results [2].

The difficulty was not the replacement but the exit. Codex was never downloadable and could not be self-hosted; access existed only through the API, so when the API closed the model was simply gone, with no way for anyone outside OpenAI to preserve a copy [2].

That hit published research hardest: studies built on Codex could no longer be reproduced or extended, because the object they measured had ceased to exist [2]. After the reaction, Sam Altman said researchers would keep access — "We didn't realize how much people liked this model; we will continue to support it for researchers!" [2].

Simon Willison, writing the day after the announcement, called it a major strategic misstep, arguing that the damage was less to any one project than to confidence in OpenAI as a stable platform to build on, at a moment when competitors and open-weight alternatives were arriving [3].

OpenAI's deprecation policy today promises minimum notice before a model is retired — at least six months for generally available models, unless safety or compliance forces a faster timeline. Codex was a free beta rather than a generally available product, and it got three days [1].

The retirement itself was defensible, and the honest version of this case says so. Codex was a free research beta that had run over a year past its welcome, its replacement was genuinely better at the task, and a vendor that can never retire anything ships nothing. OpenAI also publishes a deprecations page with dates and replacements — most vendors do not, and that page is why this case can be sourced to the record rather than to recollection [1][2].

Everything downstream had been built on the assumption that the substrate would still be there next week. The people affected had made a reasonable-looking decision — use the best available model — and an unexamined one: they had no copy, no equivalent alternative they had actually tested, and no idea what notice they were entitled to, because they had never asked. Free was the tell they missed. Nothing was breached and nothing was inaccurate; the thing simply stopped, and reproducing a published result became impossible for reasons that had nothing to do with the research [2][3].