API
OpenAI
OpenAI has deprecated a family of older audio, realtime, and speech-to-text API models — including gpt-realtime, gpt-audio, and the gpt-4o audio/realtime line — and will remove them from the API on 20 January 2027
departs 20 Jan 2027
What this means for you
On 20 July 2026 OpenAI marked a group of its voice and speech models as deprecated, and set 20 January 2027 as the date they stop responding in the API. The affected models are the ones many voice apps call directly: gpt-realtime and gpt-realtime-mini (the low-latency speech-to-speech models behind live voice agents), gpt-audio and gpt-audio-mini (audio-in/audio-out chat), the older gpt-4o-audio, gpt-4o-realtime, gpt-4o-mini-audio and gpt-4o-mini-realtime snapshots, and the gpt-4o-mini-transcribe-2025-03-20 transcription snapshot. Deprecated means they still work for now; on the shutdown date, calls that name one of these models start failing. Anything that hardcodes one of these model IDs — a phone-support bot, a live-captioning feature, a voice assistant, a transcription pipeline — will error on 20 January 2027 unless the model string is updated before then. Nothing changes on the day you read this; the six-month window is the runway to swap identifiers and re-test.
Where to go
OpenAI publishes a direct replacement for each retired model, so this is a rename rather than a re-architecture in most cases. gpt-realtime and gpt-4o-realtime map to gpt-realtime-2.1; gpt-realtime-mini and gpt-4o-mini-realtime map to gpt-realtime-2.1-mini; gpt-audio, gpt-4o-audio, gpt-audio-mini and gpt-4o-mini-audio all map to gpt-audio-1.5; and gpt-4o-mini-transcribe-2025-03-20 maps to gpt-4o-mini-transcribe-2025-12-15. Find every place your code pins one of the old model IDs (search your codebase and any config/prompt store for the exact strings), point it at the replacement, and re-run your voice/transcription test suite — newer model versions can shift latency, voice characteristics, and token accounting even when the API shape is unchanged. Teams that would rather not chase OpenAI's model cadence can evaluate other realtime-speech and speech-to-text providers, but that is a larger migration than swapping a string.
- gpt-realtime-2.1 / gpt-realtime-2.1-mini — OpenAI's named realtime replacements
- gpt-audio-1.5 — OpenAI's named audio replacement
- gpt-4o-mini-transcribe-2025-12-15 — the current transcription snapshot
If your product talks or listens, there is a good chance it calls one of the OpenAI models on this list. On 20 July 2026 OpenAI deprecated a whole cohort of its older audio, realtime, and transcription models, and gave them a hard removal date of 20 January 2027 — after which naming one of them in an API call returns an error instead of a response.
The retired models are the workhorses of the current voice-app generation: gpt-realtime and gpt-realtime-mini, the speech-to-speech models that power live voice agents; gpt-audio and gpt-audio-mini for audio-in/audio-out chat; the earlier gpt-4o-audio, gpt-4o-realtime, gpt-4o-mini-audio and gpt-4o-mini-realtime snapshots; and the gpt-4o-mini-transcribe-2025-03-20 transcription snapshot. Deprecation is a warning, not an outage — every one of these keeps working through the second half of 2026. The change that matters is dated: on 20 January 2027 they stop answering.
This is the failure mode that catches teams off guard, because nothing degrades gradually. A model ID sits hardcoded in a voice bot or a transcription job, works perfectly for months, and then throws on a single morning because the string it names no longer exists. The defense is boring and effective: treat the model identifier as a dependency with an expiry date, not a constant.
The good news is that OpenAI has published a one-to-one replacement for each retired model, so for most teams this is a find-and-replace rather than a rebuild. The realtime models move to gpt-realtime-2.1 (or gpt-realtime-2.1-mini), the audio models consolidate onto gpt-audio-1.5, and the transcription snapshot moves to gpt-4o-mini-transcribe-2025-12-15. The work is to locate every place an old ID is pinned — application code, config, prompt stores, infrastructure-as-code — repoint it, and re-run your voice and transcription tests, because a newer model version can quietly change latency, voice, and token costs even when the request shape is identical. Six months is a comfortable runway for that; the trap is only for the teams who let the date arrive first.
Entry changelog
- — last reviewed.
- — Entry created from OpenAI's official API deprecations page. Announcement dated 20 July 2026; shutdown 20 January 2027. Surfaced via Layer-2 discovery, written from OpenAI's primary deprecations list rather than a secondary report.
Spot an error or a changed date? Tell us and it’s corrected the same night.