Skip to content
TrustList
News

OpenAI API shutdowns fall on 28 September, 1 October and 23 October

Editorial

By TrustList Editorial

Four older OpenAI models stop on 28 September, GPT-5.4-Cyber on 1 October and a dozen legacy snapshots on 23 October. An inventory of model names now avoids broken integrations.

About OpenAI API shutdowns fall on 28 September, 1 October and 23 October

OpenAI API shutdowns fall on 28 September, 1 October and 23 October

24 September 2026 — OpenAI's API deprecations page lists three shutdown dates in the next five weeks. After each one, requests that name the retired models will stop working. Some of these notices are old. The one added this month is for GPT-5.4-Cyber, dated 11 September 2026. Together they are a practical deadline for anyone who still names older models in code or relies on suppliers who do.

What changes

28 September 2026. Four models announced for removal on 26 September 2025 are shut down: gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106. OpenAI's suggested replacement for all four is gpt-5.6-terra.

1 October 2026. In an entry dated 11 September 2026, OpenAI says the gpt-5.4-cyber model "is deprecated and will be removed from the API on October 1, 2026". It recommends gpt-5.6-cyber, its current security-research model, which it offers only to approved users.

23 October 2026. An entry dated 22 April 2026 lists twelve older snapshots. They include gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-turbo, gpt-4o-2024-05-13, gpt-4.1-nano, o1, o1-pro, o3-mini, o4-mini and gpt-image-1. Five fine-tuned model bases are removed on the same day. OpenAI suggests gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna or gpt-image-2 as substitutes.

Later dates are also on the page. The Evals platform becomes read-only on 31 October. Evals, reusable prompts and Agent Builder shut down on 30 November. On 26 February 2027, whisper-1 and the three gpt-4o transcription models are removed, in favour of gpt-transcribe or gpt-live-transcribe.

Who is affected

Any organisation that calls the OpenAI API directly with these model names is affected, including internal tools and scripts that were written some time ago and forgotten. Buyers of software with OpenAI features built in are affected too, because a supplier's feature may be tied to one of these snapshots. Fine-tuned models built on the retired bases stop working with them, so teams that trained custom models in 2024 or 2025 should check first.

What to do

  • Search code, configuration, prompt libraries and environment variables for the model names above. Pinned snapshot names, which include a date, are the ones most likely to be missed.
  • Test the suggested replacements on real tasks. Newer models can change output length, formatting, cost per request and refusal behaviour, so re-check anything that parses responses automatically.
  • For fine-tuned models, plan a retrain on a supported base or a switch to prompting a current model. Budget time for evaluation.
  • Ask software suppliers with OpenAI-powered features to confirm in writing which models they use and when they will move.
  • If you use Whisper for call or meeting transcription, put the February 2027 date on the roadmap now. Word timestamps, translation and speaker labels may behave differently on the new models.

None of these changes needs a contract amendment, but each is a hard stop. Requests to a retired model name fail rather than moving to a newer one.

Not yet independently verified. This rests on OpenAI’s own deprecations page; each entry there carries its own announcement date, and only the 1 October shutdown was announced in September. No independent report of it was found, so it is only as reliable as that page. We will update this when it can be confirmed, and remove this note.

Sources