Manu Shukla

Posted on Jul 22 • Originally published at ecorpit.com

18 OpenAI models shut down on 23 July 2026: the migration map, and 3 the docs page leaves out

#openai

18 OpenAI models shut down on 23 July 2026: the migration map, and 3 the docs page leaves out

Summary. Today, 23 July 2026, OpenAI shuts down a batch of model snapshots announced on 22 April 2026. The email OpenAI sent developers listed 18 snapshots. The public deprecations page currently lists 15 rows for that date. The three in the email and not in the table are all audio: gpt-4o-audio-preview-2024-12-17, gpt-4o-mini-audio-preview-2024-12-17 and gpt-4o-mini-realtime-preview-2024-12-17. Four more dates follow: 10 August 2026 for two chat-latest snapshots, 26 August 2026 for the Assistants API, 23 October 2026 for the largest wave including gpt-3.5-turbo, gpt-4, gpt-4-turbo and o4-mini, and 11 December 2026 for the original GPT-5 and o3 snapshots. One thing you may have read is not on OpenAI's page anywhere: a 20 January 2027 removal of legacy audio, realtime and transcription families. We checked the page today and it is not there.

What actually shuts down today

OpenAI announced this batch on 22 April 2026 with the standard framing: "To improve reliability and make it easier for developers to choose the right models, we are deprecating a set of older OpenAI models."

These are the 15 rows on the deprecations page for 23 July 2026, with OpenAI's own recommended substitute.

Model snapshot	Substitute OpenAI recommends	What it was for
`computer-use-preview-2025-03-11`	`gpt-5.4-mini`	Computer use
`gpt-4o-search-preview-2025-03-11`	`gpt-5.4-mini`	Web search on chat completions
`gpt-4o-mini-search-preview-2025-03-11`	`gpt-5.4-mini`	Web search on chat completions
`gpt-4o-mini-tts-2025-03-20`	`gpt-4o-mini-tts-2025-12-15`	Text to speech
`gpt-audio-mini-2025-10-06`	`gpt-audio-1.5`	Audio in, audio out
`gpt-realtime-mini-2025-10-06`	`gpt-realtime-mini`	Realtime voice
`gpt-5-chat-latest`	`gpt-5.5`	Chat
`gpt-5.1-chat-latest`	`gpt-5.5`	Chat
`gpt-5-codex`	`gpt-5.5`	Coding
`gpt-5.1-codex`	`gpt-5.5`	Coding
`gpt-5.1-codex-max`	`gpt-5.5`	Coding
`gpt-5.1-codex-mini`	`gpt-5.4-mini`	Coding
`gpt-5.2-codex`	`gpt-5.5`	Coding
`o3-deep-research-2025-06-26`	`gpt-5.5-pro`	Deep research
`o4-mini-deep-research-2025-06-26`	`gpt-5.5-pro`	Deep research

The 22 April 2026 email, reposted in full on the OpenAI Developer Forum by community manager VeitB the same day, carries three snapshots the table does not: gpt-4o-audio-preview-2024-12-17, gpt-4o-mini-audio-preview-2024-12-17 and gpt-4o-mini-realtime-preview-2024-12-17.

An earlier September 2025 notice already removed the un-suffixed aliases gpt-4o-audio-preview, gpt-4o-mini-audio-preview and gpt-4o-mini-realtime-preview on 7 May 2026, pointing them at gpt-audio-1.5 and gpt-realtime-mini. So if your code pinned the 2024-12-17 snapshot explicitly rather than the alias, today is the day it stops resolving, and the docs table will not tell you that. Grep your repository for 2024-12-17 before you read any further.

The three places OpenAI's list contradicts itself

This matters more than the individual model IDs, because it changes how much you can trust the page as a single source of truth.

The email and the table disagree on scope. Eighteen snapshots by email, fifteen in the table, and the difference is entirely audio and realtime preview snapshots. Neither list is annotated to explain the gap.

Entries have moved on and off the list. On 15 May 2026 a developer noted on the forum that gpt-4o-mini-tts-2025-03-20 had disappeared from the deprecations page and asked whether it was staying. The following day VeitB replied: "It's back! Must have been an error." A model's deprecation status flickering for a day is a reason to pin your own copy of the list rather than reading the live page at release time.

At least one recommended substitute does not do the job. A long-standing forum contributor pointed out that gpt-4o-search-preview-2025-03-11, which existed to give web search on chat completions where there are no internal tools, is documented as migrating to gpt-5.4-mini, and argued the correct destination is gpt-5-search-api. Read OpenAI's substitute column as a hint, not an instruction.

The TTS migration in particular has a quality complaint attached. Multiple forum threads in 2026 report that the recommended replacement gpt-4o-mini-tts-2025-12-15 truncates final sentences and stopped honouring the instructions parameter, which is why developers pushed to keep the March snapshot alive. If you run TTS in production, test sentence-final audio specifically before you cut over.

The claim you should not act on

Coverage circulating in July 2026 states that OpenAI notified developers on 20 July 2026 of the deprecation of legacy audio, realtime and transcription model families, with removal from the API on 20 January 2027.

We read the full deprecations page on 23 July 2026. There is no 20 July 2026 announcement on it. The most recent announcement block is dated 11 June 2026, covering GPT-5 and o3 snapshots. The only 2027 date anywhere on the page is 6 January 2027, and it concerns fine-tuning job creation, not audio.

There is also no transcription deprecation of any kind on that page: no whisper-1, no gpt-4o-transcribe, no gpt-4o-mini-transcribe, and no deprecation of /v1/audio/transcriptions. If you are planning a transcription migration on the strength of a January 2027 deadline, you are planning against a date OpenAI has not published.

Two caveats, stated plainly. OpenAI notifies affected customers by email, and email can precede the page: that is exactly what the three missing audio snapshots above demonstrate. And a page can be updated after we read it. The correct action is to check your own organisation's deprecation emails and the page itself, not to trust a secondary summary. That is the whole lesson of this article.

The next four dates

Shutdown date	What goes	Migrate to
10 August 2026	`gpt-5.2-chat-latest`, `gpt-5.3-chat-latest`	`gpt-5.5`
26 August 2026	Assistants API	Responses API and Conversations API
24 September 2026	Videos API, `sora-2`, `sora-2-pro` and their snapshots	No replacement listed
28 September 2026	`gpt-3.5-turbo-instruct`, `babbage-002`, `davinci-002`, `gpt-3.5-turbo-1106`	`gpt-5.4-mini` or `gpt-5-mini`
23 October 2026	`gpt-3.5-turbo-0125`, `gpt-4-0613`, `gpt-4-turbo`, `gpt-4.1-nano`, `gpt-4o-2024-05-13`, `gpt-image-1`, `o1`, `o1-pro`, `o3-mini`, `o4-mini`, plus fine-tuned versions	Model-specific, see the page
30 November 2026	Reusable prompts (`v1/prompts`), the Evals platform, Agent Builder	Application code, Promptfoo, Agents SDK or ChatGPT Workspace Agents
1 December 2026	`gpt-image-1-mini`, `gpt-image-1.5`, `chatgpt-image-latest`	`gpt-image-2`
11 December 2026	`gpt-5-2025-08-07`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5-pro` and `o3` snapshots	`gpt-5.5`, `gpt-5.4-mini`, `gpt-5.4-nano`, `gpt-5.5-pro`

The 23 October wave is the one that will hurt most teams, because it takes out the entire gpt-3.5-turbo and gpt-4 lines, the original gpt-4o snapshot, and the fine-tuned models built on them. Fine-tuned versions of gpt-3.5-turbo, gpt-4, gpt-4.1-nano-2025-04-14, babbage-002, davinci-002 and o4-mini-2025-04-16 all stop on the same day. If you invested in fine-tuning on those bases, the migration is a retraining project, not a string swap. That is a quarter of work, not a sprint.

The Assistants API date is separate and older: OpenAI announced on 26 August 2025 that the Assistants API would be removed one year later, on 26 August 2026, replaced by the Responses API and the Conversations API. Our Assistants API shutdown and Responses migration guide covers that path in detail.

Read the notice-period policy before you pin a model

OpenAI publishes the guarantee it holds itself to, and it is the single most useful paragraph on the page for architecture decisions.

Generally available models get at least 6 months' notice. Specialized variants of GA models get at least 3 months, and OpenAI names the categories: "chat variants such as gpt-5.1-chat-latest, Codex variants such as gpt-5.3-codex, and deep research variants such as o3-deep-research."

Preview models are the trap. OpenAI's wording: "Preview models, identified by preview in the model name, may be retired with much shorter notice, such as 2 weeks. Examples include computer-use-preview and gpt-4o-audio-preview. We don't recommend using preview models for business-critical production workloads unless you can migrate on short notice."

Three of today's shutdowns are preview models. That is the policy working as documented, not a surprise.

There is also a commercial escape hatch most teams never ask about. OpenAI states that "in some cases, developers may be able to provision dedicated capacity for continued access after a model's shutdown date" and points to its sales team. If a shutdown genuinely breaks a regulated workflow you cannot re-validate in time, that conversation is worth having before the date, not after.

What migration actually costs, in tokens

The line change is free. The bill is not, because audio pricing is duration-encoded and the substitute models sit in different price bands.

Third-party pricing trackers listed gpt-realtime-1.5 audio at $32 per million input tokens and $64 per million output tokens as of July 2026, with GPT-5.5 at $5.00 input and $30.00 output per million tokens. Audio tokens are billed by duration rather than by word, so a minute of user speech and a minute of synthesised reply are not the same number of tokens. Verify both numbers against your own OpenAI pricing page before you put them in a business case, because aggregator sites drift and OpenAI's own page is the only authority on what you will actually be charged.

The practical consequence for voice teams: a realtime migration changes latency, token accounting and cost at the same time, and you cannot infer the new bill from the old one by ratio. Run a week of shadow traffic. Our guide to building on gpt-realtime for voice agents covers the measurement setup, and the hybrid routing framework covers what to do when one provider's deprecation cadence becomes a business risk.

A migration checklist that survives the next batch

Treat deprecation as a standing engineering process, not an incident.

Grep the whole repository for dated snapshot IDs, not just aliases. The pattern to search is a four-digit year inside a model string, for example -2024-12-17 or -2025-03-11. Aliases and pinned snapshots have different shutdown dates.
Keep your own machine-readable copy of the deprecation list, checked into version control, with the date you copied it. The live page has changed retroactively at least once.
Route every model call through one adapter with the model ID in configuration. If a model ID appears in more than one file, the next batch costs you a week you did not budget.
Alert on model_not_found in production, not just in logs. On a shutdown day the failure is a 404, and it looks like a bug rather than a policy.
Subscribe a shared alias, not an individual, to OpenAI's deprecation emails. The email is the earliest and most complete signal, and today proves it is more complete than the docs page.
Test the substitute against your own evaluation set before the date, especially for audio. The TTS complaints above are a live example of a recommended replacement that changed behaviour.
For fine-tuned models on the 23 October list, start now. Retraining, re-evaluating and re-approving a fine-tune is measured in weeks.

India-specific considerations

Indian teams carry two extra constraints on the same calendar.

Voice and transcription workloads in India are often the ones with the tightest cost ceilings, because they support vernacular support flows priced against Indian ARPU rather than US ARPU. A realtime model migration that raises per-minute cost by a third is a product decision in Bengaluru, not just an infrastructure one, and it usually forces a fresh look at routing cheaper turns to a smaller model.

The compliance side is quieter but real. If your transcription or voice pipeline processes Indian personal data, swapping the model swaps the processor configuration behind a system your DPDP records describe. Update the record of processing when you update the model ID, not at the next audit. Teams that have already run one of these migrations end to end will recognise the pattern from our LLM migration and cost optimisation work.

The models are getting better and the shutdown cadence is getting faster. Both facts point at the same engineering answer: stop pinning model IDs in application code.

FAQ

Which OpenAI models shut down on 23 July 2026?

OpenAI's deprecations page lists fifteen snapshots for that date. They include the computer use preview, both of the older search preview snapshots, one text to speech snapshot, one audio mini snapshot, one realtime mini snapshot, five Codex snapshots, two chat latest snapshots, and the two deep research snapshots built on o3 and o4 mini.

Why does OpenAI's email list more models than the docs page?

The 22 April 2026 email, reposted in full on the OpenAI Developer Forum, named eighteen snapshots for 23 July 2026, while the public page lists only fifteen. The three absent from the table are the December 2024 audio preview, mini audio preview and mini realtime preview snapshots. OpenAI has not annotated the difference anywhere on the page.

Is OpenAI deprecating Whisper or its transcription models?

Not according to the deprecations page as read on 23 July 2026. No whisper-1, gpt-4o-transcribe or gpt-4o-mini-transcribe entry appears on it, and the transcriptions endpoint is not listed. Reports of a January 2027 transcription cutoff do not match what OpenAI has published.

When does the Assistants API shut down?

On 26 August 2026. OpenAI notified developers on 26 August 2025 that the Assistants API would be removed one year later, and points migrations at the Responses API together with the Conversations API, using its Assistants to Conversations migration guide.

What is the biggest deprecation date still ahead?

23 October 2026. It removes the whole GPT-3.5 turbo line, GPT-4 and GPT-4 turbo, GPT-4.1 nano, the original May 2024 GPT-4o snapshot, the first GPT image model, and the o1, o1 pro, o3 mini and o4 mini reasoning models, along with fine-tuned versions built on several of those bases. That makes it a retraining project rather than a config change.

How much notice does OpenAI promise before retiring a model?

At least 6 months for generally available models and at least 3 months for specialized variants such as chat, Codex and deep research variants. Preview models, identified by preview in the name, may be retired with as little as 2 weeks' notice, and OpenAI advises against using them for business-critical production workloads.

Can I keep using a model after its shutdown date?

OpenAI says that in some cases developers may be able to provision dedicated capacity for continued access after a model's shutdown date, and directs those requests to its sales team. That is an exception to negotiate before the date, not a default, and it does not appear as a self-serve option.

What breaks first when a model is shut down?

The API returns a model-not-found error, so calls fail rather than degrade. Teams that pin dated snapshots in application code discover this as a production 404 on the shutdown day, which is why routing every call through one configurable adapter is the single highest-value change to make.

How eCorpIT can help

eCorp Information Technologies Private Limited is a Gurugram-based, CMMI Level 5 assessed technology consultancy founded in 2021, and our senior engineering teams run model migrations as a normal part of keeping AI systems in production. That means an adapter layer with model IDs in configuration, an evaluation set that runs against the substitute before the cutover, and cost modelling done on your own traffic rather than on list prices. If you have fine-tuned models on the 23 October 2026 list, the retraining window is already short. Start at /contact-us/, or read how we scope LLM migration and cost optimisation.