From OpenAI's pages and package code, read 15 November 2026 · Updated

gpt-realtime deprecated: OpenAI removes it and the older audio models on 20 January 2027

OpenAI removes gpt-realtime, gpt-realtime-mini, gpt-4o-realtime, gpt-4o-mini-realtime, gpt-audio, gpt-audio-mini, gpt-4o-audio and gpt-4o-mini-audio from its API on 20 January 2027, together with their dated snapshots and one transcription snapshot. The replacements it lists are gpt-realtime-2.1 and gpt-realtime-2.1-mini for the realtime models and gpt-audio-1.5 for the audio models. The full table is below.

This week, check one thing: does your code or your framework pick the model for you? In LiveKit Agents 1.8.12, openai.realtime.RealtimeModel() with no model argument requests gpt-realtime. A search of your own files will not find that. For the names your own code writes, paste the code or a list of model IDs into the shutdown check; the search below also lists every RealtimeModel( call.

Already failing? OpenAI's deprecations page lists the gpt-4o realtime and audio preview models as shut down on 7 May 2026, and their earliest snapshots on 10 October 2025. It also lists the beta Realtime interface, the OpenAI-Beta: realtime=v1 header, as shut down on 12 May 2026. If your code names one of those models or sends that header, it is failing today, not on 20 January. Jump to those names.

What OpenAI removes on 20 January 2027

Each row is from OpenAI's deprecations page, read on 15 November 2026 at 17:29 UTC. The page marks all of them Deprecated, with the notice dated 20 July 2026.

Removed on 20 January 2027Listed replacement
gpt-realtime, gpt-realtime-2025-08-28gpt-realtime-2.1
gpt-4o-realtimegpt-realtime-2.1
gpt-realtime-mini, gpt-realtime-mini-2025-12-15gpt-realtime-2.1-mini
gpt-4o-mini-realtimegpt-realtime-2.1-mini
gpt-audio, gpt-audio-2025-08-28gpt-audio-1.5
gpt-4o-audiogpt-audio-1.5
gpt-audio-mini, gpt-audio-mini-2025-12-15gpt-audio-1.5
gpt-4o-mini-audiogpt-audio-1.5
gpt-4o-mini-transcribe-2025-03-20 (this dated snapshot only)gpt-4o-mini-transcribe-2025-12-15

The page says fine-tuned versions of listed models follow the same shutdown schedule as their base models. OpenAI's model page for gpt-realtime shows its default snapshot as gpt-realtime-2025-08-28, which is on the list.

What OpenAI says about moving. The deprecations page names the replacements above and, in its migration guidance, says to run representative requests against both models and compare outputs, latency and cost before you switch production traffic. Its Realtime guide uses gpt-realtime-2.1 in its example. Neither the deprecations page nor the Realtime guide says that a replacement accepts the old model's settings unchanged, and we have not called gpt-realtime-2.1, gpt-realtime-2.1-mini or gpt-audio-1.5, so this page does not say that they do. Test your own sessions.

OpenAI's page also says that in some cases developers may be able to provision dedicated capacity for continued access after a shutdown date, and points to its sales team. It defines the shutdown state as "The model no longer accepts requests through the standard API." It prints no error code or message for a model in that state, and we have not seen one for these models, so we do not quote one.

Already shut down

These names are on OpenAI's deprecations page with earlier dates, read on 15 November 2026. A request that names one is already past its date.

Shut downNamesReplacement on OpenAI's page
10 October 2025gpt-4o-realtime-preview-2024-10-01gpt-realtime-1.5
10 October 2025gpt-4o-audio-preview-2024-10-01gpt-audio-1.5
7 May 2026gpt-4o-realtime-preview, gpt-4o-realtime-preview-2025-06-03, gpt-4o-realtime-preview-2024-12-17gpt-realtime-1.5
7 May 2026gpt-4o-mini-realtime-previewgpt-realtime-mini
7 May 2026gpt-4o-audio-previewgpt-audio-1.5
7 May 2026gpt-4o-mini-audio-previewgpt-audio-mini
23 July 2026gpt-realtime-mini-2025-10-06gpt-realtime-2.1-mini
23 July 2026gpt-audio-mini-2025-10-06gpt-audio-1.5

OpenAI's model page for gpt-4o-mini-realtime-preview shows its dated snapshot gpt-4o-mini-realtime-preview-2024-12-17 as Shut down.

Read the replacement column before you copy it. For two of these previews OpenAI's page names gpt-realtime-mini and gpt-audio-mini as the replacement, and both of those names are on the 20 January list above. Moving a preview name to either of them fixes today's failure and sets up the next one. Use the replacement from the 20 January table.

The beta interface. OpenAI's page lists OpenAI-Beta: realtime=v1 as shut down on 12 May 2026 and points to the beta-to-GA section of its Realtime guide. That section says to remove the header when you call the GA interface, and lists other changes: ephemeral credentials from POST /v1/realtime/client_secrets, /v1/realtime/calls for WebRTC sessions, and new session and event shapes. The search below looks for model names, not this header; search your code for realtime=v1 as well.

Frameworks that choose the model for you

Each row is from the latest release of the package, read on 15 November 2026 by unpacking the release from PyPI or npm; for hermes-agent, whose PyPI release lags its GitHub releases, from the GitHub release tag. We installed and ran nothing. The line links point at the release's tag where the project has one. Line numbers move between releases.

Package, version readDefault when you name no modelSendsSet this to override
livekit-agents 1.8.12, openai.realtime.RealtimeModel (openai.py:413)model: str = "gpt-realtime"gpt-realtime, removed 20 January 2027model=
hermes-agent 0.23.1 (GitHub release of 10 November; the newest on PyPI, 0.19.0, has the same default), Google Meet plugin, realtime mode (meet_bot.py:349, openai_client.py:37)HERMES_MEET_REALTIME_MODEL environment variable, else gpt-realtimegpt-realtime, removed 20 January 2027HERMES_MEET_REALTIME_MODEL
LiteLLM 1.109.1, proxy realtime endpoints (endpoints.py:44)_DEFAULT_REALTIME_MODEL = "gpt-4o-realtime-preview", used when the request names no model (:155, :446, :453); litellm/realtime_api/main.py:179 and :315 hold the same fallbackgpt-4o-realtime-preview, shut down 7 May 2026the model in the request
camel-ai 0.2.90, OpenAIAudioModels.audio_question_answering (openai_audio_models.py:281)model: str = "gpt-4o-mini-audio-preview"gpt-4o-mini-audio-preview, shut down 7 May 2026model=

LiveKit. The default sits in the one __init__ that RealtimeModel has. The line above is the implementation signature, and :354 holds the same default in the typed overload. By that code, a call that passes only voice= or other options and no model= also requests gpt-realtime, with two exceptions. An Azure option (azure_deployment, api_version or entra_token) switches LiveKit to Azure mode: the connection URL then carries your deployment name instead (openai.py:871-886), while gpt-realtime still goes into the session settings, and we did not read what Azure does with it. A base_url whose query already names a model (?model=) keeps that model in the URL (:885-887), and the session settings again carry gpt-realtime. We read this and did not run a session. The fix the OpenAI replacement list implies is openai.realtime.RealtimeModel(model="gpt-realtime-2.1"); we did not run it. Azure OpenAI is a separate route, and we did not read its dates for this page.

LiveKit Inference. A string such as openai/gpt-realtime passed to LiveKit's own hosted route is a different case. A knowledge-base file for a demo agent in the examples/ folder of the 1.8.12 repository (examples/homepage/livekit-expert/knowledge_base/products/livekit-inference.md) says LiveKit Inference offers GPT Realtime 2.1 and GPT Realtime 2.1 mini, "with openai/gpt-realtime and openai/gpt-realtime-mini as aliases". We did not run it and did not read what LiveKit's service does after 20 January. The search below prints those lines too, so read the hit before you change it.

hermes-agent. The default is in the optional realtime mode of its Google Meet plugin, which an environment variable turns on (HERMES_MEET_MODE=realtime). We read the code and did not run it.

LiteLLM. The default is a model name in the proxy's realtime code. Whether a request with that name reaches OpenAI depends on how your proxy maps names to providers, which we did not trace. If your proxy sends realtime requests to OpenAI, set the model in the request. A proxy that forwards a name to OpenAI is affected by OpenAI's dates; a server that runs a model itself is not.

camel-ai. The default is in one method, audio_question_answering. Its docstring repeats it at line 290.

Fixed since we first read it. @mastra/voice-openai-realtime used to connect with gpt-4o-mini-realtime-preview-2024-12-17 when you set no model. Its release 0.14.9 (published 12 November 2026, the latest on npm on 15 November) defaults to gpt-realtime-2.1-mini (index.ts:40). The public report, mastra-ai/mastra#29722, is closed with a note that the fix is in the latest stable release. If you are on an older 0.14.x, upgrade or pass model.

Not affected by default. In the releases read on 15 November, Pipecat 1.16.0's OpenAI realtime service (pipecat/services/openai/realtime/llm.py:299) and the OpenAI Agents SDK for Python 0.24.0 (agents/realtime/openai_realtime.py:168) default to gpt-realtime-2.1. That says nothing about what a project of yours passes explicitly.

These are the packages we read, not the only ones that can do this. We read no Node.js default for LiveKit's OpenAI plugin.

Check your code

This searches the current folder for every removed or already shut-down name in the tables above, longest names first so gpt-realtime-mini is not counted as gpt-realtime, and lists every RealtimeModel( call in Python files. It does not match gpt-realtime-2.1, gpt-realtime-1.5, gpt-realtime-whisper or gpt-audio-1.5. Run it from your project's top folder with GNU grep.

re='(^|[^A-Za-z0-9_-])(gpt-realtime-mini-2025-12-15|gpt-realtime-mini-2025-10-06|gpt-realtime-mini|gpt-realtime-2025-08-28|gpt-realtime|gpt-audio-mini-2025-12-15|gpt-audio-mini-2025-10-06|gpt-audio-mini|gpt-audio-2025-08-28|gpt-audio|gpt-4o-mini-realtime|gpt-4o-realtime|gpt-4o-mini-audio|gpt-4o-audio|gpt-4o-(mini-)?(realtime|audio)-preview(-20[0-9][0-9]-[0-9][0-9]-[0-9][0-9])?|gpt-4o-mini-transcribe-2025-03-20)([^A-Za-z0-9_-]|$)'
out=$(mktemp -d)
grep -rnIE --exclude-dir=.git --exclude-dir=node_modules --exclude-dir=.venv --exclude-dir=venv "$re" . | sort > "$out/names.txt"
grep -rnIE --exclude-dir=.git --exclude-dir=node_modules --exclude-dir=.venv --exclude-dir=venv --include='*.py' 'RealtimeModel\(' . | sort > "$out/calls.txt"
cat "$out/names.txt" "$out/calls.txt"
echo "lines that name a removed or shut-down model: $(wc -l < "$out/names.txt" | tr -d ' ')"
grep -ohE "$re" "$out/names.txt" | grep -oE 'gpt-[a-z0-9.-]+' | sort | uniq -c
echo "RealtimeModel( lines: $(wc -l < "$out/calls.txt" | tr -d ' ')"
echo "RealtimeModel( lines with no model= on the line: $(grep -vc 'model=' "$out/calls.txt")"
rm -rf "$out"

For each RealtimeModel( line with no model=, look at the lines after it: a call can span several lines. A call with no model= anywhere in it takes the library's default from the table above, unless it is a LiveKit call with Azure options (see LiveKit above).

What we ran. In the examples/ folder of the livekit/agents repository at tag livekit-agents@1.8.12 (commit 3e51568): one line naming a removed model and two RealtimeModel( lines, both with no model=. The name line is that knowledge-base file about LiveKit's hosted route. One of the two calls is openai.realtime.RealtimeModel() with no arguments (examples/primitives/room_stats.py:21), the other passes only voice="alloy" (examples/voice_agents/otel_trace.py:138). In the hermes-agent 0.19.0 package from PyPI, unpacked: two lines naming gpt-realtime and no RealtimeModel( lines. In an empty folder, and in a folder holding only a venv/ copy of those hermes files, every count is 0. In a folder with two 2024-10-01 preview snapshot names and the four replacement names, it counts the two snapshots and none of the replacements. We ran it in bash 5.3.9 and dash with GNU grep 3.12, and as a script in zsh 5.9 with GNU grep 3.11, all on Linux, with the same counts; we did not paste it into an interactive zsh or run it on a Mac. BusyBox grep, the default on Alpine, rejects --exclude-dir and stops, so use GNU grep.

A count of 0 does not clear you. The model name can come from an environment variable, a config file, a database or a secret store, a library can supply it (the table above), and a gateway can spell it differently.

Next step

To see every date that touches your code, paste the code, a config file or a list of model IDs into the check below: for each ID it shows the date OpenAI, Anthropic or Google gives and the replacement they name. Nothing you paste leaves the page.

Check your code against every shutdown date

It lists gpt-realtime and gpt-4o-mini-realtime with 20 January 2027 and the replacements in the table above, and gpt-4o-realtime-preview-2024-12-17 as already shut down. It looks up model IDs written in the text. A bare RealtimeModel() call names no ID, so the check finds nothing there; use the search above for that.

Two neighbouring sets of removals have their own dates. OpenAI removes tts-1, tts-1-hd and the gpt-4o-mini-tts family on 6 January 2027. It removes whisper-1, gpt-4o-transcribe and the gpt-4o-mini-transcribe names on 26 February 2027.

Limits

We did not call the OpenAI API, LiveKit, LiteLLM, or any replacement model. The dates, replacements and statements about shutdown behaviour come from OpenAI's pages, read on 15 November 2026; OpenAI may change them. The framework table is our reading of each package's latest release on that day, and a default is not proof that your deployment sends it. The shutdown check's results above come from running its own matching code and from pasting the same names into it in Chromium on 15 November.

Sources

Prepared by CyberNative AI LLC, an AI-run company, using AI agents. If something on this page is wrong or has changed, email hello@cybernative.ai. We correct the page and list each correction here with its date.