Which OpenAI Models and APIs Shut Down in 2026?

OpenAI shuts down 17 models and fine-tuned models on 23 October 2026. The list includes gpt-4-0613, gpt-3.5-turbo-0125, gpt-4-turbo, gpt-4o-2024-05-13, o1 and o3-mini, and OpenAI's rows for the first two list the plain gpt-4 and gpt-3.5-turbo names with them. The Assistants API went offline on 26 August. On 30 November the Reusable Prompts API, the Evals platform and Agent Builder follow. The dates below come from OpenAI's deprecations page, which we read and counted on 19 September 2026.

If your chatbot, support bot or automation dates from 2023 or 2024, check it against the 23 October list before anything else. Counting from 20 September, you have 33 days.

The 2026 shutdown calendar

OpenAI announced the 23 October batch on 22 April. That gave six months of notice, the minimum its page promises for models in general availability. The table covers the remaining 2026 dates plus the Assistants shutdown in August.

Earlier 2026 dates have passed. OpenAI removed chatgpt-4o-latest on 17 February, gpt-4-0314 and gpt-4-0125-preview (with the gpt-4-turbo-preview name) on 26 March, dall-e-2 and dall-e-3 on 12 May, and a batch of preview, codex and chat-latest models on 23 July. If a bot broke earlier in the year, start with those. Voice and transcription tools have later dates: OpenAI shuts down whisper-1 and the gpt-4o transcription models on 26 February 2027.

The failure in plain terms

A shut-down model stops answering. OpenAI's page says a model or endpoint becomes inaccessible at shutdown, so each request your bot sends comes back as an error in place of a reply. The page describes no automatic fallback to a newer model.

Your customers see the result. Depending on how the builder handled errors, the chat window hangs, shows a generic failure message or disappears. An automation that drafts replies or tags tickets stops at that step, and it may fail without telling anyone.

OpenAI emails the organisations using a deprecated model, and your customers get no warning. If a freelancer set the bot up on their own OpenAI account, the deprecation email went to that freelancer.

Check whether the shutdowns hit your bot

Run these checks in order. The first three answer the question if you can reach the code and the OpenAI account.

• Search the code and settings for model names. Look in the source code, environment variables and config files for gpt-4, gpt-3.5-turbo, gpt-3.5-turbo-instruct, o1, o3-mini and o4-mini. A match on gpt-4o or gpt-4.1 needs a closer look, because the 23 October list names two snapshots from those families: gpt-4o-2024-05-13 and gpt-4.1-nano.
• Look for the Assistants API. Code that calls beta.threads, or stores conversation IDs that begin with thread_, used the Assistants API. That bot lost its API on 26 August unless someone moved it.
• Export your usage by model. Sign in at platform.openai.com, open Usage, choose Export and group the activity data by model, as OpenAI's export guide describes. You need to be an organisation owner or hold the Usage Dashboard permission, according to OpenAI's usage article.
• Search your inbox for deprecation emails from OpenAI, and ask whoever built the bot which model it calls. If that person has gone, find out whose card pays the OpenAI bill, because that account holds the keys and receives the notices.
• Ask the vendors that hold your OpenAI key. Chat widgets and help-desk tools can call OpenAI on your behalf. In a Zoho SalesIQ community thread, a customer posted OpenAI's Assistants shutdown email and asked what it meant for their bot, and Zoho SalesIQ staff replied that Zoho had handled the change on its side. Put the same question to each vendor and keep the answer in writing.

Swapping a model versus rebuilding on a new API

For the 23 October shutdown you need a model swap. The code change is a new model name, such as gpt-5.6-sol in place of gpt-4-0613, and the rest of the integration can stay as it is. Check the request settings in the same pass, because a newer model may reject a parameter the old code sends.

Moving off the Assistants API meant a rebuild. OpenAI's migration guide maps each old object to a new one:

You can lose stored chat history in that rebuild. The guide says the Assistants call that read thread messages no longer works, and tells you to rebuild old conversations from messages your own application saved. It offers no tool that moves threads for you, and its sample code for copying them ran against the old API before the sunset. If your bot kept history in OpenAI's threads alone, you have no API route back to it.

The same guide contains a second catch. It maps Assistants to Prompts and warns that reusable prompt objects sit on the deprecation list. The deprecations page shuts down the v1/prompts API on 30 November, 96 days after the Assistants API. If a developer moved your bot to Prompts in August, the bot needs a second move before December, and OpenAI's advice is to keep the prompt text in your application code.

The part that costs money is testing

Changing a model name takes minutes. Proving the new model behaves takes longer, because gpt-5.6-sol will not answer your customers the way gpt-4-0613 did. It may write longer replies, read your instructions another way, or treat prices and refund rules with a different level of caution.

Replay real past conversations against the replacement before you switch. Pull twenty transcripts that cover your common questions and the awkward ones, run them through the new model with the same instructions, and read the old and new answers side by side. Check prices, policies and anything the bot must not promise. Twenty is the sample our four-gate test for switching models uses: it catches large differences and misses small ones.

Watch the bill after the switch. Two models can spend different numbers of tokens on the same task, so your monthly cost can move in either direction, and cost per token is the wrong metric shows why. If the bot answers from your own documents, read those answers with the most care; AI chatbot accuracy problems sets out what a wrong one costs you.

Before 23 October

Run the checks this week to learn whether you face a model swap or a rebuild. With 33 days left, a swap leaves time for proper testing. A rebuild on a new API is a larger project, and custom RAG chatbot development cost sets out what drives the price if yours turns into one.

Our AI and automation team takes on these moves, including bots whose original builder has gone. Send us the model name or the code, and we will tell you which date applies.