On this page · 11 sections
- What OpenAI is switching off, and when
- The catch: the TTS replacement is not a Speech API model
- What breaks on Vapi, Retell, LiveKit and Pipecat
- Migration table: old model to replacement
- Latency and cost considerations
- Worked cost example: 10,000 four-minute support calls a month
- Test plan for telephony voice agents
- India: data residency, TRAI and DPDP
- Migration timeline
- How eCorpIT can help
- References
Summary. OpenAI removes every Speech API model, tts-1, tts-1-hd and both gpt-4o-mini-tts snapshots, on 6 January 2027 and names gpt-realtime-2.1-mini as the replacement. That is not a model swap. The replacement does not run on /v1/audio/speech; it runs only on the Realtime API, so narration jobs, IVR prompt generators and chained speech-to-text, LLM and text-to-speech pipelines need new code. On 20 January OpenAI also removes the first-generation gpt-realtime, gpt-realtime-mini, gpt-audio and gpt-audio-mini models and their gpt-4o aliases, and on 26 February whisper-1 and the gpt-4o transcription models. On price, tts-1 costs $15 per million characters, about $0.0135 a spoken minute, while gpt-realtime-2.1-mini charges $20 per million audio output tokens, about $0.024 a minute at one output token per 50 ms. Realtime 2.1 keeps the old $32 and $64 audio rates but raises text output from $16 to $24. On 1 November 2026 Vapi's OpenAI voice provider still offered only the retiring TTS models, and LiveKit's OpenAI realtime plugin still defaulted to gpt-realtime. Pin model IDs, test voices and barge-in on real phone audio, and finish before year-end change freezes.
About this article. This guide is for teams running voice agents, IVR and narration on OpenAI audio models, directly or through Vapi, Retell, LiveKit or Pipecat. Shutdown dates, prices, endpoint support, voices and platform defaults were checked on 1 November 2026 against OpenAI's deprecations page, model pages, pricing page and audio guides, Microsoft Learn's retirement schedule, the four platforms' documentation and price lists, and TRAI's September 2026 press release. Platform defaults change without notice, so re-check before you commit; this is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.
What OpenAI is switching off, and when
OpenAI's deprecations page carries three notices that hit voice products in the first quarter of 2027: text-to-speech (announced 1 October 2026), legacy audio and Realtime models (announced 20 July 2026) and transcription (announced 26 August 2026).
| Shutdown date | Models or endpoint removed | OpenAI's recommended replacement | Code change needed |
|---|---|---|---|
| 6 January 2027 | tts-1, tts-1-hd | gpt-realtime-2.1-mini | Yes: Speech API to Realtime API |
| 6 January 2027 | gpt-4o-mini-tts (2025-03-20 and 2025-12-15 snapshots) | gpt-realtime-2.1-mini | Yes: Speech API to Realtime API |
| 20 January 2027 | gpt-realtime (2025-08-28), gpt-4o-realtime | gpt-realtime-2.1 | Model ID, then session review |
| 20 January 2027 | gpt-realtime-mini (2025-12-15), gpt-4o-mini-realtime | gpt-realtime-2.1-mini | Model ID, then session review |
| 20 January 2027 | gpt-audio (2025-08-28), gpt-audio-mini (2025-12-15), gpt-4o-audio, gpt-4o-mini-audio | gpt-audio-1.5 | Model ID on Chat Completions |
| 20 January 2027 | gpt-4o-mini-transcribe-2025-03-20 | gpt-4o-mini-transcribe-2025-12-15 | Snapshot pin, good only to 26 February |
| 26 February 2027 | whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-transcribe-diarize | gpt-live-transcribe or gpt-transcribe | Model ID, sometimes endpoint |
| 26 February 2027 | /v1/audio/translations | v1/realtime/translations | New endpoint |
OpenAI's policy allows at least six months' notice for generally available models and at least three months for specialised variants. The TTS notice says it comes with at least three months' notice without saying which category OpenAI applied. The same page says some developers may be able to arrange dedicated capacity through sales to keep a model after shutdown; treat that as a negotiation for large accounts, not a plan.
Text models retire in the same window. GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini leave on 14 January 2027, so a chained agent with gpt-4o-mini as its brain faces two migrations eight days apart; see our GPT-4o and GPT-4.1 shutdown guide.
Azure OpenAI runs its own calendar. Microsoft's schedule retires tts 001, tts-hd 001 and whisper 001 on 15 December 2026, three weeks before OpenAI's TTS date, while gpt-realtime 2025-08-28 runs to 2 March 2027 and gpt-4o-mini-tts 2025-12-15 to 15 June 2027.
The catch: the TTS replacement is not a Speech API model
OpenAI's text-to-speech guide now opens with a notice that the move to gpt-realtime-2.1-mini changes the API workflow: the new model is not supported on /v1/audio/speech, and existing Speech API integrations keep working only until the shutdown. The model page agrees. Of the 18 endpoints it lists for gpt-realtime-2.1-mini, only v1/realtime is supported; Batch, Chat Completions, Responses and Speech generation are all marked unsupported.
That breaks narration jobs that send text and save an MP3, pre-rendered IVR menus and disclosures, chained speech-to-text, LLM and text-to-speech agents, and platform assistants that pick OpenAI as the TTS provider. OpenAI's own Agents SDK voice pipeline is one of them: its default TTS model on the main branch is gpt-4o-mini-tts.
You have four exits, and most teams will mix them:
- Realtime API in text-in, audio-out mode. Open a gpt-realtime-2.1-mini session over WebSocket, WebRTC or SIP, send text and receive audio. Sessions handle 24 kHz PCM and G.711 μ-law (
audio/pcmu), and the voice cannot change once the session has produced audio.
- gpt-audio-1.5 on Chat Completions. A request-response shape at $2.50 and $10 per million text tokens and $32 and $64 per million audio tokens. If its output audio is counted like Realtime's, at one token per 50 ms, speech costs about $0.077 a minute. Use it for small, stateless jobs.
- A third-party TTS. LiveKit Inference's Build and Ship price list shows Inworld Realtime TTS 2.0 Flash, Fish Audio S2.1 Pro and AssemblyAI TTS 1 at $15 per million characters, the same rate as tts-1; Gemini 3.8 Flash TTS at $24, Deepgram Aura-2 at $30 and Cartesia Sonic 3 at $50. For Indian languages, LiveKit documents a Sarvam TTS plugin using
bulbul:v3withhi-INand other Indic targets, plus an 8 kHz telephony sample rate.
- GPT-Live for the whole conversation. OpenAI's full-duplex voice model, generally available since 10 September 2026 at $0.05 per session minute billed per second, replaces the full loop rather than only TTS. It suits open conversation, not scripted wording, as covered below. Our GPT-Live build guide goes deeper.
Voices do not map one to one. tts-1 and tts-1-hd support alloy, ash, coral, echo, fable, onyx, nova, sage and shimmer. The Realtime list is alloy, ash, ballad, coral, echo, sage, shimmer, verse, marin and cedar, with marin or cedar recommended for quality. Brands built on fable, onyx or nova have to re-cast, and anyone steering tone with gpt-4o-mini-tts instructions must move that steering into session instructions.
Exact wording is a real risk. Vapi's GPT-Live migration guide says exact or prerecorded speech is not available because GPT-Live generates all of its speech, including the greeting. LiveKit's GPT-Live plugin has no text-only mode, so it cannot be paired with a separate TTS. Realtime models generate speech too, so check that consent notices, price quotes and regulatory disclosures come out word for word, or keep pre-recorded audio for them.
What breaks on Vapi, Retell, LiveKit and Pipecat
We checked each platform's documentation on 1 November 2026, and none had published a note naming the 6 January or 20 January cutoffs. Defaults can change before then, so read this as today's risk. For pricing and feature differences between the platforms, see our Vapi, Retell, Bland and LiveKit comparison.
| Platform | Exposure on 6 January (TTS) | Exposure on 20 January (Realtime) | What to do now |
|---|---|---|---|
| Vapi | OpenAI voice provider accepts only tts-1, tts-1-hd and gpt-4o-mini-tts | Realtime picker offers gpt-realtime-2025-08-28 and gpt-realtime-mini-2025-12-15 | Move OpenAI voices to another provider or a Realtime model; switch realtime assistants to gpt-realtime-2 or GPT-Live |
| Retell | OpenAI is one of its TTS voice providers; no OpenAI TTS notice yet | Its April 2026 swap put gpt-4o-mini-realtime agents on gpt-realtime-mini, which retires on 20 January | Confirm the plan with Retell; also handle its 31 January 2027 model-field change |
| LiveKit | OpenAI TTS plugin defaults to gpt-4o-mini-tts | OpenAI plugin RealtimeModel defaults to gpt-realtime | Set the model explicitly; LiveKit Inference already offers only gpt-realtime-2.1, its mini and GPT-Live |
| Pipecat | OpenAITTSService defaults to gpt-4o-mini-tts | OpenAI Realtime service defaults to gpt-realtime-2.1 | Change TTS model or provider; move input transcription off gpt-4o-transcribe before 26 February |
Vapi. Its Realtime page labels gpt-realtime-2025-08-28 as GPT Realtime Cluster for existing assistants and recommends gpt-realtime-2 for new ones. Its GPT-Live migration guide, marked beta, lists what does not carry over: other voice providers and cloned voices, squads and handoffs, model and voice fallbacks, exact or prerecorded speech, and knowledge bases; transfers are blind only.
Retell. Its changelog says GPT Realtime 2.1 is live in the agent builder and lists OpenAI among its TTS voice providers. On 3 April 2026 Retell replaced gpt-4o-realtime with gpt-realtime-1.5 and gpt-4o-mini-realtime with gpt-realtime-mini, and that mini model is on OpenAI's 20 January list. Retell also retires legacy request fields such as model and s2s_model in favour of model_config on 31 January 2027, a second change in the same month for API users. Its deprecation page has an RSS feed, which is where a notice would appear.
LiveKit. The OpenAI realtime plugin's parameter table gives gpt-realtime as the default model, and its own example passes only a voice. Code written as openai.realtime.RealtimeModel(voice='marin') will point at a retired model on 20 January unless the default changes first. The OpenAI TTS plugin defaults to gpt-4o-mini-tts.
Pipecat. OpenAITTSService defaults to gpt-4o-mini-tts, while the OpenAI Realtime service already defaults to gpt-realtime-2.1 and lists gpt-4o-transcribe among its input transcription options.
Migration table: old model to replacement
Prices are per million tokens unless stated, from OpenAI's model pages and pricing page.
| Old model | Replacement | Old price | New price | Interface change | Watch for |
|---|---|---|---|---|---|
| tts-1 | gpt-realtime-2.1-mini | $15 per 1M characters | $20 per 1M audio output tokens; $0.60 per 1M text input | Speech API to Realtime session | fable, onyx and nova voices gone; verbatim reading |
| tts-1-hd | gpt-realtime-2.1-mini | $30 per 1M characters | As above | As above | Quality A/B on phone audio |
| gpt-4o-mini-tts | gpt-realtime-2.1-mini | $0.60 text in, $12 audio out | As above | As above | Tone instructions move into the session |
| gpt-realtime | gpt-realtime-2.1 | $32/$64 audio; $4/$16 text | $32/$64 audio; $4/$24 text | Model ID; reasoning settings | Reasoning tokens; context grows from 32,000 to 128,000 |
| gpt-realtime-mini | gpt-realtime-2.1-mini | $0.60/$2.40 text | $0.60/$2.40 text; $10/$20 audio | Model ID | Reasoning effort per turn |
| gpt-audio-mini | gpt-audio-1.5 | $0.60/$2.40 text | $2.50/$10 text; $32/$64 audio | Model ID | Higher cost; consider Realtime mini |
| whisper-1 | gpt-transcribe or gpt-live-transcribe | $0.006 a minute | $0.0045 or $0.017 a minute | Model ID; live needs a streaming session | Language hints for code-switching |
The legacy gpt-realtime-mini and gpt-audio-mini pages list only text rates, so compare audio spend from your own usage logs. For the transcription side, our gpt-transcribe and gpt-live-transcribe comparison covers accuracy and cost.
Latency and cost considerations
Turn audio into minutes. OpenAI's voice cost guide counts user audio at one token per 100 ms (600 a minute) and assistant audio at one token per 50 ms (1,200 a minute). Applied to the published rates, that gives:
| Model | Caller audio in, per minute | Agent audio out, per minute | Cached audio, per 1M tokens |
|---|---|---|---|
| gpt-realtime-2.1 and legacy gpt-realtime | $0.0192 | $0.0768 | $0.40 |
| gpt-realtime-2.1-mini | $0.0060 | $0.0240 | $0.30 |
| gpt-live-1 | $0.05 per session minute, billed per second, plus backend model costs | Included in the session rate | Not applicable |
Replay is the hidden cost. Every Realtime response re-sends the whole conversation, and the guide says caching is best effort with no guarantee. Keep instructions and tools stable at the top of the session. When truncation kicks in, the server default retention_ratio of 1.0 drops only what it must and busts the cache on every turn; a lower value such as 0.8 drops extra history and keeps more turns cacheable.
Reasoning adds latency and tokens. gpt-realtime-2.1 supports configurable reasoning effort, and OpenAI warns that higher effort can raise latency and output token use. Teams moving from legacy gpt-realtime, which had no reasoning setting, should start at a low effort and raise it only for turns that make policy decisions. Our gpt-realtime-2.1 build guide covers the settings.
GPT-Live trades token maths for a clock. It bills active session time, including silence and time the backend spends working, and creating a WebRTC session bills 15 seconds up front that are credited once the session runs. OpenAI's launch post reports 0.798-second turn-taking for GPT-Live-1 against 1.41 seconds for GPT-Realtime-2.1, and, with GPT-6 Astra at medium effort as the backend, 83.6% first-attempt completion on the Tau3 support benchmark against 45.7%; both are OpenAI's own tests. Early adopters in the same forum thread report that the split between voice model and backend makes scripted flows and clean hang-ups harder.
Chained to Realtime removes a hop but gives up the text checkpoint between what the agent decided and what it said. Keep that checkpoint if you run policy checks on replies before they are voiced.
Worked cost example: 10,000 four-minute support calls a month
Assumptions. 10,000 inbound calls a month, 4 minutes each (40,000 minutes), 10 turns per call. In each turn the caller speaks for 12 seconds (120 audio tokens) and the agent for 9.6 seconds (192 tokens), so each call has 2 minutes of caller audio and 1,920 agent audio output tokens. Each response re-reads the conversation so far. We assume the previous turn's whole input is served from cache, while the agent's last reply and the caller's new audio are billed fresh: across 10 turns that is 2,928 fresh and 12,312 cached audio input tokens. Instructions and tools add 2,000 text tokens per turn (20,000 a call, 18,000 of them cached), plus 1,000 text output tokens a call for reasoning and text. Telephony and platform fees are excluded. Prices come from OpenAI's pricing and model pages.
gpt-realtime-2.1-mini per call: fresh audio 2,928 x $10/1M = $0.0293; cached audio 12,312 x $0.30/1M = $0.0037; audio output 1,920 x $20/1M = $0.0384; text input 2,000 x $0.60/1M plus 18,000 x $0.06/1M = $0.0023; text output 1,000 x $2.40/1M = $0.0024. Total $0.0761 a call, or $0.0190 a minute.
| Stack | Per call | Per minute | Per month (10,000 calls) |
|---|---|---|---|
| Legacy gpt-realtime (retiring) | $0.2527 | $0.0632 | $2,527 |
| gpt-realtime-2.1 | $0.2607 | $0.0652 | $2,607 |
| gpt-realtime-2.1-mini | $0.0761 | $0.0190 | $761 |
| gpt-live-1 with gpt-6-luna backend | $0.2027 | $0.0507 | $2,027 |
| gpt-live-1 with gpt-6.1-terra backend | $0.2270 | $0.0568 | $2,270 |
The GPT-Live rows are 4 minutes x $0.05 plus six delegations per call (18,000 input and 1,800 output tokens) at $0.10 and $0.50 per million for gpt-6-luna, or $1 and $5 for gpt-6.1-terra. Realtime 2.1 sits $80 a month above legacy only because of the higher text output price (10 million tokens x $8 per million); heavy reasoning widens the gap.
If caching misses, every replayed token is billed at the fresh rate: the mini rises to about $0.205 a call ($0.051 a minute) and the full model to about $0.715 a call ($0.179 a minute). Log cached_tokens from each response.done event in week one.
For a chained agent that only swaps TTS, 1.6 minutes of agent speech per call at about 900 characters a minute is 14.4 million characters a month, or $216 on tts-1. Moving that speech to gpt-realtime-2.1-mini costs about $384 in audio output (19.2 million tokens x $20 per million), a $15-per-million-character third-party voice stays at $216, and Deepgram Aura-2 at $30 comes to $432.
Test plan for telephony voice agents
Run this on a staging number with real carrier audio, not a laptop microphone. Record every call, keep the old configuration live, and compare. The audio-path row follows OpenAI's SIP guide.
| Area | What to run | Pass mark |
|---|---|---|
| Inventory | Search code, infrastructure as code and platform configs for tts-1, gpt-4o-mini-tts, gpt-realtime, gpt-audio, whisper-1 and unpinned defaults | Every call site has an explicit new model ID |
| Audio path | Inbound and outbound over your SIP trunk with G.711 μ-law at 8 kHz; OpenAI SIP needs TLS on port 5061 and SRTP to four published address ranges | No one-way audio, no clipping, DTMF passes |
| Barge-in | Interrupt the agent at the start, middle and end of a long answer, 20 times each | Agent stops within your target and does not repeat the cut-off text |
| Turn-taking | Long pauses, filler sounds, hold music, a TV in the background, a second speaker | No premature replies; no talking over the caller |
| Voices | Blind A/B of old versus new voice with 10 listeners on phone audio | New voice preferred or equal on clarity and brand fit |
| Verbatim scripts | Consent notice, recording disclosure, price and refund lines, 50 runs each | 100% word for word, or route to pre-recorded audio |
| Alphanumerics | Order IDs, PIN codes, +91 numbers, rupee amounts, dates read back | Correct read-back on an agreed sample |
| Languages | English with Indian accents, Hindi, Hinglish and each regional language you serve | Intelligibility and intent accuracy at your bar |
| Tools and hang-up | Every tool path, failed tool, transfer and end of call | End-call tool fires; no duplicate bookings |
| Transcripts | Confirm the input transcription model and that both sides of the transcript are stored | Analytics, QA and CRM fields unchanged |
| Latency | p50 and p95 time to first audio, per region and carrier | Within 10% of today, or an agreed new target |
| Cost | Per-call tokens or seconds from usage events for 500 calls | Within 15% of the worked estimate |
Indian languages need explicit coverage. OpenAI's TTS guide says the model broadly follows Whisper's language support, which includes Hindi, Marathi, Kannada, Tamil, Urdu and Nepali, and notes that the voices are tuned for English. Bengali, Telugu, Gujarati, Malayalam and Punjabi are not on that list. On 16 October 2026 OpenAI added a Hindi voice, Nira, for GPT-Live. A 15 October announcement on OpenAI's developer forum said GPT-Live-1.1, with better Hindi among other languages, was rolling out in ChatGPT, Codex and dots and would reach the API later; the API changelog showed no release by 1 November. Route weak languages to an Indic TTS such as Sarvam or a BHASHINI-based stack.
India: data residency, TRAI and DPDP
OpenAI's data controls guide lists /v1/realtime and /v1/live/sessions for regional storage and processing in the United States and Europe (EEA and Switzerland) only. Its India region, in.api.openai.com, offers regional storage but not regional processing, requires Modified Abuse Monitoring or Zero Data Retention, and does not list the Realtime or GPT-Live endpoints. For DPDP planning, treat call audio sent to Realtime or GPT-Live as processed outside India, say so in your notice, and set retention limits on recordings and transcripts.
Outbound AI calls also fall under TRAI's Third Amendment to the commercial communications regulations, announced on 18 September 2026. It defines application-to-person calls to include robo-calls and pre-recorded or artificial voice, requires every entity using them to declare that use and its calling numbers to its telecom operator in advance, treats undeclared calls as unsolicited commercial communication, and adds a termination charge on such calls. If your consent notice must be read word for word, test it as a verbatim script.
Migration timeline
- By 15 November 2026: inventory every model ID and platform default, and decide per use case between Realtime mini, gpt-audio-1.5, a third-party TTS or GPT-Live.
- By 1 December: build the new path in staging, pin explicit model IDs and run the test plan on carrier audio.
- By 15 December: canary 10% of live traffic, compare cost and latency, and fix voice or script issues before holiday freezes. Azure users must also be off tts, tts-hd and whisper 001 by this date.
- Before 6 January 2027: move all Speech API traffic and remove tts-1 and gpt-4o-mini-tts from code and configs.
- Before 20 January: move all gpt-realtime, gpt-realtime-mini and gpt-audio traffic to 2.1 models or GPT-Live.
- Before 26 February: move transcription and translation off whisper-1, the gpt-4o transcribe models and
/v1/audio/translations.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our AI voice agent development service runs migrations like this one, from model and platform inventory through Realtime or GPT-Live rebuilds, Indic TTS integration and SIP testing on Indian carriers to cost tracking from usage logs. Talk to us at /contact-us/.
Last updated: 1 November 2026.
