OpenAI shuts off older GPT-3.5 and GPT-4 API models on 23 October

What happened

OpenAI stopped accepting requests for a group of older GPT-3.5 and GPT-4 model names on 23 October 2026. According to OpenAI's developer deprecations page, the list includes snapshots such as gpt-3.5-turbo-0125 and gpt-4-0613, along with several GPT-4 Turbo, GPT-4o, o1 and o4-mini names. The same page says the fine-tuned versions built on the retired base models were shut down on that date as well.

OpenAI first announced the retirement on 22 April 2026, about six months before the shutdown. The deprecations page lists a named substitute for each retired name. Most GPT-3.5 names point to gpt-5.6-terra, and most GPT-4 names point to gpt-5.6-sol.

An independent AI news brief covering the same day, Deep Gains, reported that OpenAI retired seventeen legacy API model IDs, including GPT-4 and o1. The count depends on how names are grouped. A developer write-up on the shutdown lists 17 model IDs, while a monthly tracker counts 28 IDs when aliases and fine-tuned names are included. The OpenAI page is the reference for which names are affected, so check it rather than a count.

Why it matters for businesses that automate work

Many automations do not use a single product screen. They call a model by its name, sometimes deep inside a workflow tool, a spreadsheet add-on or a script that a former employee wrote. Invoice extraction, email drafting, report summaries and ticket triage are common examples. When the name is retired, the request fails, and the failure may only show up when the next batch of work arrives.

A developer write-up on the shutdown makes the point plainly: once a model is retired, requests to its ID can fail, and a replacement should not be assumed to behave the same way. That is the core risk for a business. The workflow still runs, but it may produce different wording, different field values or a different format.

What changes for a business that automates work, in practice

The change is a configuration task first and a testing task second. The steps below suit a five-person office and a large group alike. The size of the list is the only difference.

1. List every place a model name is stored. This includes code, no-code workflow tools, saved prompts, configuration files, scheduled jobs and vendor settings that someone else manages for you.
2. Mark the fine-tuned models separately. The OpenAI page treats them as retired with their base model, so they need to be rebuilt on a supported base.
3. Pick the replacement named on the page for each name, then run the new setting on a copy of the workflow using real past examples.
4. Compare the output with the old results. Check amounts, dates, names, tone and any fields that other systems read automatically.
5. Switch the live workflow only after the comparison passes, and keep a dated record of what was changed and who approved it.

The table below shows how several retired names map to the replacements listed on OpenAI's page at the time of writing.

Some businesses will find that a name is not listed where they expected. A name that is an alias for a retired snapshot may still need attention, so the safest approach is to check the exact string in each system and not rely on memory.

Where automation cannot help

Automation cannot judge whether a new model's answers are good enough for a given task. A person still needs to read samples from invoices, customer replies or reports before the switch goes live. Output from a newer model can differ in wording, length and structure, even when the request is identical. Where a workflow sends figures to tax, payroll or accounting systems, the person responsible for that data should approve the test results before any live change.

What to watch next

The deprecations page is the place to check. It also lists upcoming items. According to the same page, a group of text-to-speech models will be removed from the API on 6 January 2027, with at least three months' notice, and the page recommends gpt-realtime-2.1-mini as the migration target. Businesses that use voice features should add that date to their calendar now.

OpenAI's page also lists the retirement of older image models in a separate entry. Any business that uses image generation through the API should check that entry, because the page and third-party trackers do not always agree on dates or names. Confirm each date on the official page before planning a change.