Realtime API migration: what changes and who is affected

If your business has anything that listens or talks, a phone line that answers itself, a voice feature in an app, a transcription step buried in a support workflow, then it is calling a named OpenAI model. That name has a shutdown date attached to it, and on the voice surface those dates have been arriving roughly twice a year.

There is a live one. On 20 Jul 2026 OpenAI announced the deprecation of its legacy audio, realtime and transcription model families, with shutdown on 20 Jan 2027. This post establishes who that affects, what the previous rounds already broke, and how to arrange things so the next round is a configuration change rather than a project.

Every date and model name below was checked against OpenAI's own deprecations page and model documentation on 08 Aug 2026. OpenAI's docs move on roughly a monthly cadence, so re-check before you act on anything here.

Four rounds, three of which have already happened

The voice surface has churned harder than the text surface. Laid out in order, the pattern is the point.

Read the first row and the last row together, because that is the finding.

In September 2025 OpenAI told developers to move off the preview models and onto four named replacements. Two of those four, gpt-realtime-mini and gpt-audio-mini, are in the July 2026 deprecation list. A business that did exactly what the vendor advised in April 2026 is in the next round anyway, about ten weeks later.

That is not a criticism of OpenAI. It is the nature of a capability that is still improving quickly. But it does settle the planning question. If you treat each of these as a one-off disruption to be absorbed, you will absorb one every few months indefinitely. The useful response is structural, and it is cheap.

Establishing whether you are affected

Three checks. None of them needs a developer for longer than an afternoon.

Find the model string. Somewhere in your system, in configuration or in code, is a line naming the model. Have someone search for realtime, gpt-audio, transcribe and 4o-audio across the repository and the environment variables. Whatever comes back, compare it against the fourth row of the table above. If it matches, you have a date.

Check whether the string is pinned or floating. This matters more than it sounds. On 13 Jan 2026 OpenAI repointed the unversioned gpt-realtime-mini and gpt-audio-mini slugs at newer snapshots, leaving the previous versions reachable only by their dated identifiers. So an unversioned slug changes the model underneath you without anybody deploying anything, and a pinned dated snapshot does not change but goes dark on a published date. Both are defensible. Neither is safe if nobody knows which one you chose.

Check what your system actually did on 7 and 12 May 2026. If your voice feature has been quietly broken since May 2026, or degraded in a way that got written off as flakiness, the two shutdowns in that week are the first place to look. This is worth an explicit check rather than an assumption, because a voice system that fails at the model call often fails silently from the caller's point of view: they hear dead air or a fallback message, and they hang up rather than raise a ticket.

If all three come back clean, stop. Most businesses running a voice feature built in the last year are already on a current model and have nothing to do until the next announcement.

What actually changes, and what needs re-testing

The migration itself is usually a small change. The Realtime API is a distinct API with its own endpoints, and it did not move: sessions run against /v1/realtime, ephemeral browser credentials come from /v1/realtime/client_secrets, WebRTC sessions from /v1/realtime/calls, and the three transports (WebRTC, WebSocket and SIP for telephony) are unchanged. You are swapping a model name, not rebuilding an integration.

The cost is not in the swap. It is in the fact that the new model behaves differently on purpose.

OpenAI's own description of gpt-realtime-2.1 is that it updates the previous version "with improved alphanumeric recognition, silence and noise handling, and interruption behavior". Every one of those three is something you tuned around. If your agent reads back order numbers, the recognition path changed. If you set silence thresholds to stop it talking over people, the behaviour those settings were compensating for changed. Improvements are still changes, and a voice system that was tuned against the old behaviour needs re-tuning against the new one.

Two more things to check on the way through.

gpt-realtime-2.1 exposes configurable reasoning effort, and OpenAI notes that higher reasoning effort can increase latency and output token usage. On a phone call, latency is what the caller judges you on. Whoever does the migration should set that deliberately and measure the result rather than accept a default, because it trades directly against the thing your callers judge you on.

And the model carries a 128,000 token context window with a knowledge cutoff of 30 Sep 2024. If your agent answers questions about your products, pricing or policies, it is not doing that from the model's memory and never was. It is doing it from whatever your system feeds it. That plumbing is the part that determines whether the answers are right, and it is unaffected by any of these dates.

Build so the next round is a configuration change

There will be another announcement. The question is only whether it costs you a project or an afternoon. Four things make the difference, and all four are cheaper to do now than under a deadline.

Put the model name in configuration, not in code. If changing model requires a code change, a review and a release, every deprecation is a development ticket. If it is an environment variable, it is a decision followed by a restart. This is the single highest-return item on the list and it is usually an hour of work.

Keep a set of recorded calls you can replay. Twenty or thirty real interactions covering your common intents, your awkward ones, a caller who interrupts, a caller reading out a long reference number, a caller with background noise. Without this, "does the new model still work" is answered by a colleague trying it twice and saying it seemed fine. With it, you get a real answer in an hour, and you get the same answer every time a model changes. This is the asset, more than the code.

Subscribe someone to the deprecations page. Not a team, a named person, with a diary entry to look at it monthly. Every date in this post was published in advance. The businesses that get hurt are not the ones that were not told, they are the ones where nobody was listening.

Know who owns the OpenAI account. It is often in the name of a contractor, or of someone in finance who has never heard of any of this. Deprecation notices go to that address. Sort it out before you need it.

The call in front of you

Run the three checks. If you are on the January 2027 list, you have a date and a named replacement, which is the easy version of this problem: schedule it, budget a re-tuning pass rather than just a swap, and use the enforced downtime to put the model name in configuration and build the replay set.

If you are not affected, do the four structural items anyway. They cost less than one emergency migration and they convert every future announcement from an incident into a line in a change log.

If you are earlier than that, and still deciding whether an AI voice agent belongs on your phone line at all, the migration question is the wrong one to start with. We have written separately on what a voice agent survives on a real UK phone line, covering latency, accents, call recording and escalation to a human.

If you have not yet decided which route you are on, how the OpenAI options compare for a UK business covers the ground.

We are an OpenAI Select Partner and we do this work, but we take no margin on your usage and we have no reason to turn a model swap into a rebuild. If you want your exposure established and priced, request a quote and tell us what you think is running. The same team handles OpenAI implementation and Claude work on identical terms, and if the honest answer is that you have nothing to do until January, we will tell you that. The same principle applies to the other dated shutdown in front of UK businesses in 2026, the Assistants API retirement.