OpenAI announced the deprecation of gpt-realtime on 20 July 2026 and removes it from the API on 20 January 2027 — 128 days away as at 14 September 2026. Nothing breaks today. The published replacement is gpt-realtime-2.1, and the work that decides whether you land it is re-tuning your prompt, not swapping a model string.
“We’re in production on gpt-realtime and it’s being deprecated” — what the notice says
OpenAI’s deprecations page, read on 14 September 2026, carries an entry headed 2026-07-20: Legacy audio, realtime, and transcription models. It states that on 20 July 2026 OpenAI notified developers using legacy audio, realtime and transcription model families and snapshots of their deprecation and removal from the API on 20 January 2027. Nine rows are listed, all with the same shutdown date:
| Model family / snapshot | Recommended replacement | Shutdown date |
|---|---|---|
gpt-realtime |
gpt-realtime-2.1 |
20 Jan 2027 |
gpt-4o-realtime |
gpt-realtime-2.1 |
20 Jan 2027 |
gpt-realtime-mini |
gpt-realtime-2.1-mini |
20 Jan 2027 |
gpt-4o-mini-realtime |
gpt-realtime-2.1-mini |
20 Jan 2027 |
gpt-audio |
gpt-audio-1.5 |
20 Jan 2027 |
gpt-4o-audio |
gpt-audio-1.5 |
20 Jan 2027 |
gpt-audio-mini |
gpt-audio-1.5 |
20 Jan 2027 |
gpt-4o-mini-audio |
gpt-audio-1.5 |
20 Jan 2027 |
gpt-4o-mini-transcribe-2025-03-20 |
gpt-4o-mini-transcribe-2025-12-15 |
20 Jan 2027 |
A second entry dated 2026-08-26 covers whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, removed 26 February 2027, replaced by gpt-live-transcribe or gpt-transcribe. Name a transcription model separately and you have two migrations on two deadlines.
The same page sets the policy behind those dates: at least six months of notice for generally available models, at least three months for specialised variants, as little as two weeks for anything with preview in the name. The realtime notice is 184 days. It also records one escape hatch, verbatim — “In some cases, developers may be able to provision dedicated capacity for continued access after a model’s shutdown date” — a sales conversation rather than an API flag.
Is this as urgent as it feels?
Less urgent than it feels. Deprecation and shutdown are separate events: by OpenAI’s definitions on that page, a model is deprecated the moment it is announced, and the shutdown date is when it stops being accessible. In between it serves production traffic normally. Nothing is degrading today because of the notice.
Not urgent if you run a chained pipeline, where speech recognition, the text model and speech synthesis are separate components and you are replacing one box; your voice runs on a platform that owns the upstream account, making this one email; or your prompt is short and your tool schema small.
Urgent if your prompt was tuned against a dated snapshot and encodes strict tool-calling; your script has been through compliance review, so any behavioural change reopens it; you are on the SIP path, where webhook events and acceptance payloads belong to one API contract; or you have fine-tunes, which never outlive their base model. And plainly: if the answer to “which model IDs are in our call path” is a shrug, that is the urgent thing rather than the date, and it costs an afternoon and no money to fix.
The migration threshold table: what is still safe at N days out
Rows are keyed to 20 January 2027. Substitute your own shutdown date and the windows move with it. Options narrow as a step function, not a slope.
| Days to shutdown | Window | Still safe to start | Stop doing this |
|---|---|---|---|
| More than 120 | To 22 Sep 2026 | Change architecture: realtime model to chained pipeline, or to a delegating full-duplex model. Change SIP provider. Rewrite the prompt. | Nothing. The only window a rebuild fits in. |
| 120 to 60 | 22 Sep to 21 Nov 2026 | Swap inside the same API contract, re-split the prompt, build the regression suite, run parity against baseline calls, ship to a traffic slice. | Starting a transport or vendor change. A new call path needs a soak window you no longer have. |
| 60 to 14 | 21 Nov 2026 to 6 Jan 2027 | Prompt fixes, tool-schema tightening, staged ramp, rollback to the deprecated model while it still answers. | Changing architecture, framework version or telephony provider in the same window. One variable at a time. |
| Fewer than 14 | From 6 Jan 2027 | Running at 100% on the replacement, monitoring, old model ID wired only as an abort switch. | Any migration work. Starting here is choosing which failure to accept. |
| Zero | 20 Jan 2027 | Nothing. Requests naming retired IDs stop being served. | Assuming a grace period. The only published extension is dedicated capacity, arranged in advance. |
The Contract Line: migrate or re-architect?
The Contract Line: if the replacement keeps your handshake, your credentials and your event names, it is a migration and you budget prompt re-tuning. If any one of the three changes, it is a re-architecture and you budget a call-path rebuild.
That is the vendor’s own warning turned into a test. OpenAI’s audio guide states that sharing a transport does not make GPT-Live and Realtime handshakes, credentials or event formats interchangeable; its telephony guide adds that Realtime webhook events, call identifiers and acceptance payloads belong to the Realtime API, while a Live session uses the GPT-Live contract. Two products, one transport, incompatible contracts.
So gpt-realtime to gpt-realtime-2.1 sits below the line — same API, same events, same credentials — and the cost is behavioural, not structural. gpt-realtime to GPT-Live sits above it: different handshake, different webhook event, and the prompt split in two. That is a rebuild, feasible at 128 days out and not at 60.
The behavioural cost below the line is documented. OpenAI’s Realtime prompting guide says to prompt Realtime 2 as a reasoning voice agent rather than a basic voice bot, lists spoken preambles as first-class behaviour before tool use, warns that conflicts between overlapping always and never rules are more costly, and records the context window growing from 32k to 128k tokens. So a prompt written for a non-reasoning model does not carry across unchanged, and the usual symptom is a spoken preamble where you expected silence — the failure family covered in our guide to why voice agents go silent on tool calls. On OpenAI’s own forum, topic 1387746 (opened 21 July 2026, now closed) records one operator measuring 98 to 100% tool-call success on the deprecated snapshot and materially lower on the newer models; in topic 1386141 (opened 9 July 2026, still open as at 14 September 2026) a reported SIP tool-calling failure was not reproduced by an OpenAI staff responder, who pointed to the prompt, conversation state or turn segmentation rather than the model, nor by another developer, who reported gpt-realtime-2.1-mini calling tools once the prompt written for the older model is adjusted. Single-team results, not benchmarks, and the signal is not that the replacement is broken — it is that prompts do not transfer.
Worked calculation: does the migration fit in the days you have?
The window ran 184 days, 20 July 2026 to 20 January 2027; as at 14 September 2026, 56 are gone and 128 remain. Substitute your own inputs.
Inputs. Six call flows, a forty-scenario regression suite, one engineer at 0.6 FTE, below the Contract Line so no transport change. These inputs and the day counts below are illustrative estimates written to be substituted, not measurements taken from any particular project.
Effort. Re-split and re-tune the prompt, 3 days. Capture baselines and build the suite, 5 days. Parity runs at 2 days per iteration, budgeted at four iterations rather than one, 8 days. Tool-schema and confirmation boundaries, 4 days. Total 20 engineering days, which at 0.6 FTE is 33 working days, about 47 calendar days. Add a staged rollout at 10%, 50% then 100% with a week of observation at each step, 21 days, and a 14-day freeze before the shutdown. Total 82 calendar days.
Slack. 128 minus 82 is 46 days, comfortable only because the suite is inside the estimate. Drop the suite and you have not saved five days, you have removed the instrument that tells you whether the migration worked.
The start-date rule that falls out. Latest safe start = shutdown date minus (engineering days ÷ FTE × 1.4, as calendar days) minus 35 days of rollout and freeze. On these inputs, 30 October 2026. If your number lands in the past you are triaging, not migrating, and the honest options are one flow instead of six, more FTE, or asking about dedicated capacity.
How do I tell whether my voice stack is exposed at all?
This is the durable question and it outlives this notice. A voice agent does four jobs: hear, decide, call tools, speak. Exposure to any model shutdown is a count of how many of those four a single model ID performs.
OpenAI’s voice-agents guide sets out three architectures and, read as a risk document, grades itself. A chained pipeline runs speech-to-text, the agent workflow, then text-to-speech, and its stated reason to exist is that each component can be replaced independently. A full-duplex live model handles the spoken interaction and delegates reasoning and tools to a backend you choose. The Realtime API uses one model to interpret audio, decide what to do and respond in speech — four jobs, one string.
The four-job exposure test takes about an hour. Grep the repository for every model identifier in the call path. Record whether each is an alias or a dated snapshot — an alias such as gpt-realtime can resolve to a snapshot like gpt-realtime-2025-08-28, and the table lists both kinds. Record which of the four jobs it does, then write its shutdown date beside it. For anything on a platform rather than your own key, ask which model IDs your agents resolve to and what notice you are given — an answer that belongs in the contract, the ground our model deprecation risk assessment for AI agents covers before you pick a vendor.
Any ID scoring four out of four is a single point of failure with a published date attached. That is not an argument against speech-to-speech — it buys turn-taking and prosody a pipeline works hard to match, and the trade is in our comparison of voice AI cost per minute, realtime versus pipeline — but it is an argument for knowing which of the two you bought.
What to do in the next 24 hours, then the next 7 days
The first four cost nothing, involve no vendor and no purchase, and can be done tonight.
- Inventory. Run the four-job exposure test. One page: every model ID, alias or snapshot, job, shutdown date.
- Find who got the email. The deprecations page states that customers actively using a model are notified by email. If nobody received it, your billing contact and your engineering team are not the same people.
- Freeze a baseline. The OpenAI migration guide asks you to save representative conversations with their starting state, expected tool actions, final application state and spoken responses. Twenty to forty of them is a workable suite size, and the baseline is what later separates a model regression from a prompt bug.
- Write down the pass mark before you see results: tool-call success rate, time to first audio, and a count of unexpected spoken preambles — the discipline in our guide to testing an AI voice agent before go-live.
Then over the following seven days:
- Apply the Contract Line and state in writing whether this is a migration or a re-architecture.
- Re-split the prompt. Conversation style stays with the voice model; workflows and tool rules move to the backend. Delete overlapping always and never rules.
- Run the suite against the replacement and publish both numbers side by side. A natural-sounding reply is not evidence that a tool ran.
- Wire the rollback before the ramp. Model selection belongs in configuration so reverting is a deploy, not a rebuild — see building failover into an AI voice agent.
- If your slack is negative, stop migrating and start negotiating: ask about provisioned dedicated capacity, in writing, with your usage attached.
Do it yourself, or hand the call path over?
The method above is complete, and someone with the flows in their head and a free fortnight will land it. The crossover, as numbers rather than advice:
| Your situation | Do it yourself | Hand the call path over |
|---|---|---|
| Distinct call flows | Up to about 3 | More than about 6, or any flow under compliance review |
| Prompt size | Under roughly 1,500 words | Long prompts with strict entity capture and conditional parameters |
| Regression suite | Exists and runs unattended | Does not exist, and building it is most of the project |
| Engineering availability | 0.5 FTE or better, uninterrupted | Below 0.3 FTE, or the same engineer owns the on-call roster |
| Days remaining at the start | More than 90 | Fewer than 60, or a second deprecation clock also running |
| Who holds the model contract | Your own API key and billing relationship | A platform account where you cannot see or pin the model ID |
The last row decides whether this is ever your problem again. Zian AI runs private model deployment on customer infrastructure, so the model identifier and its lifecycle sit inside an arrangement you can read rather than inside someone else’s account; what else carries across when you change platforms is in our guide to switching AI voice platforms and what transfers. Zian AI is in partnership-application beta, which is the honest scope of what we can claim here.
Frequently asked questions
When exactly does gpt-realtime stop working?
On 20 January 2027, per the OpenAI deprecations page as at 14 September 2026. That entry was announced on 20 July 2026 and covers nine model families and snapshots, including gpt-realtime, gpt-audio, gpt-4o-realtime, gpt-realtime-mini and gpt-audio-mini. The model continues to serve production traffic normally until that date.
Is a deprecated model already degraded or rate limited?
No. OpenAI defines deprecation as the announcement that a model is being retired, and shutdown as the date it stops being accessible. Between the two dates the model behaves as before. The OpenAI deprecations page, read on 14 September 2026, also states the minimum notice periods behind those dates: at least six months for generally available models, at least three months for specialised variants, and as little as two weeks for preview models.
Can I get an extension past the shutdown date?
The only route OpenAI publishes is provisioned dedicated capacity. The deprecations page says developers may in some cases be able to provision dedicated capacity for continued access after a shutdown date, and directs them to the sales team. It is not an API setting and it is not guaranteed, so treat it as a negotiation to open early rather than a fallback you can rely on.
Will my prompt work unchanged on the replacement model?
Assume not. The OpenAI prompting guide for Realtime models describes the newer generation as a reasoning voice agent rather than a basic voice bot, makes spoken preambles a first-class behaviour, warns that overlapping always and never rules are more costly, and records the context window expanding from 32k to 128k tokens. Re-tuning the prompt is the main body of migration work.
Does this notice affect my speech-to-text model too?
Only if you name one separately. A second entry dated 26 August 2026 covers whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize, with removal on 26 February 2027 and gpt-live-transcribe or gpt-transcribe as replacements. That is a different deadline, 37 days after the realtime one, so schedule it as a second project.
If the exposure test shows one model identifier doing all four jobs on your own key, the migration is yours and the method above is the whole of it. If it shows a platform account you cannot inspect, the next step is a conversation about where the model lifecycle sits.