Azure gpt-realtime retirement: 25 June 2027 voice-agent upgrade plan

Summary. Microsoft's retirement schedule lists gpt-realtime-2.1 and gpt-realtime-2.1-mini (version 2026-07-07) retiring on Azure OpenAI on 25 June 2027, 353 days after their version date and well short of the 18-month life Microsoft's policy sets for GA models. The 2025-12-15 TTS, transcription and audio-mini models go ten days earlier. No replacement is listed; Microsoft names one about 90 to 120 days before retirement, so between 25 February and 27 March 2027. OpenAI, by contrast, lists no shutdown for gpt-realtime-2.1 on its own API. On Azure, gpt-realtime-2.1 costs $32 per million audio input tokens and $64 per million audio output tokens; the mini costs $10 and $20. The alternatives are gpt-live-1 at $3.00 per hour of voice session plus backend tokens, the managed Voice Live API, or gpt-realtime-1.5, which lasts only until 24 August 2027. Realtime API SIP runs only in East US 2 and Sweden Central. Set NoAutoUpgrade, retest phone audio, barge-in, tool calls and p95 time to first audio on every transport, and finish by mid-May.

About this article. For teams running voice agents on Azure OpenAI in Microsoft Foundry. On 7 November 2026 we checked dates against Microsoft Learn's retirement schedule (last updated 23 September 2026) and lifecycle policy, prices against the Azure pricing page and Retail Prices API, endpoints and regions against Microsoft Learn, OpenAI's position against its own documentation, and telecom rules against TRAI's press release of 18 September 2026. Microsoft revises its schedule often. This is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

What retires in June 2027

These rows come from Microsoft's schedule and OpenAI's deprecations page. None lists a replacement.

Earlier Azure dates: tts, tts-hd and whisper (001) on 15 December 2026, gpt-realtime-2 (Preview) on 20 January 2027 with gpt-realtime-2.1 as replacement, gpt-realtime (2025-08-28) on 2 March 2027, and gpt-realtime-mini (2025-10-06) on 6 April 2027, a version that also carries a past 21 September 2026 row while region tables still list it. OpenAI's January removals are covered in our guide to OpenAI's January 2027 TTS and Realtime shutdowns.

Three things about these dates that need a decision

The life is short. Microsoft's policy says GA models retire 18 months after launch and close to new customers at 12 months; its 12-month exception names only Anthropic, DeepSeek, Fireworks and Mistral AI models. gpt-realtime-2.1 gets under 12 months, so the Deprecated stage never happens, and Microsoft does not explain why. gpt-image-1.5 and the gpt-image-2.5 models also get 12 months. An Azure team that leaves gpt-realtime before 2 March 2027 moves again by 25 June, 115 days later. Budget for yearly audio migrations.

The duplicate rows conflict. gpt-realtime-mini 2025-12-15 appears with both 15 June 2027 and 15 December 2026. Plan on the earlier date and check the per-SKU deprecationDate your subscription sees in the Models API.

Preview deployments land on a short runway. Microsoft force-upgrades retiring previews with at least 30 days' notice, so anyone on gpt-realtime-2 lands on gpt-realtime-2.1 on 20 January 2027 with about five months left.

Upgrade paths available on Azure today

Microsoft tells customers to evaluate other models before a replacement is named. Five paths exist today, priced from Azure's own lists.

GPT-Live is the largest change. Microsoft describes a full-duplex voice API: the agent listens while it speaks, and a separate backend handles reasoning, lookups and tools. gpt-live-1 (version 2026-09-10) is listed for Global Standard in Canada Central, Central US, East US 2, West US 3, France Central, Sweden Central, South India and UAE North. Its Azure meter of $3.00 an hour took effect on 1 September 2026, matching the $0.05 a minute OpenAI set when GPT-Live 1 became generally available on its API on 10 September. It is not on the retirement schedule, so its lifetime is unknown.

Voice Live is the managed option. Microsoft lists gpt-realtime-2.1, gpt-realtime-2.1-datazone and gpt-realtime-2.1-regional as Pro-tier models, gpt-realtime-2.1-mini as Standard, and chained text models such as gpt-5.6-terra and gpt-6-luna. You don't deploy the models, and events mostly match the Realtime API. A SIP route in public preview covers resources in East US 2, France Central, Southeast Asia, Sweden Central, West US and West US 2. Microsoft does not say how Voice Live handles the 25 June date, so ask.

Global Standard, Data Zone Standard and Standard deployments auto-upgrade at retirement, region by region, to the listed replacement where supported; provisioned ones never do. An unset versionUpgradeOption behaves like OnceCurrentVersionExpired; NoAutoUpgrade never upgrades and stops at retirement. An unplanned swap can change voice, turn-taking and tool behaviour on live calls, so set NoAutoUpgrade and cut over on your own date.

Realtime API checks before any model swap

Most breakage on a model change comes from the session layer.

• Use the GA endpoint. Microsoft says to use /openai/v1 for all Realtime models and not to send date-based API versions or the api-version query parameter. Move code still on a preview path first, in its own release.

• Write the GA event shape. OpenAI removed its beta Realtime interface (OpenAI-Beta: realtime=v1) on 12 May 2026. In the GA shape, session.update carries session.type set to realtime, output_modalities set to audio or text but not both, and format objects such as audio/pcm at 24,000 Hz. Some Azure samples still use beta-era input_audio_format and modalities; don't copy them.

• Fix the transcription dependency. Azure's one documented deviation from OpenAI's spec is that input_audio_transcription.model takes your deployment name. If that deployment runs gpt-4o-mini-transcribe 2025-12-15, it retires ten days before the voice model. Microsoft names gpt-live-transcribe as its current recommended realtime transcription model.

• Re-check token limits. Microsoft gives 32,000 input and 4,096 output tokens; OpenAI lists a 128,000-token context and 32,000 output tokens. Test long calls on Azure.

• Handle the 60-minute cap. Azure Realtime sessions last at most 60 minutes. Read expires_at from session.created and renew with saved context.

• Tune reasoning effort. OpenAI says gpt-realtime-2.1 supports configurable reasoning effort and that more effort can raise latency and output tokens.

A GA-style session update that sends G.711 mu-law audio to a phone leg, following OpenAI's documented example:

            {
  "type": "session.update",
  "session": {
    "type": "realtime",
    "output_modalities": ["audio"],
    "instructions": "You are the billing support agent. Confirm before any refund.",
    "audio": {
      "input": {
        "format": { "type": "audio/pcm", "rate": 24000 },
        "turn_detection": { "type": "semantic_vad" }
      },
      "output": {
        "format": { "type": "audio/pcmu" },
        "voice": "alloy"
      }
    },
    "tools": []
  }
}
          

WebRTC, WebSocket and SIP: what to re-test

The flows below come from Microsoft's Realtime, WebRTC, SIP and GPT-Live pages.

WebRTC detail. Microsoft recommends WebRTC for most real-time audio. A safer variant proxies the SDP negotiation through your token service so the browser never holds the ephemeral token, and a server can attach an observer WebSocket at /openai/v1/realtime?call_id=<id> to watch or record the session.

SIP detail. Realtime SIP runs only in swedencentral and eastus2. The accept call to /openai/v1/realtime/calls/{call_id}/accept requires type, model and instructions, and model is your deployment name, so every accept payload changes with the deployment. Reject sends SIP 603 by default, refer takes a tel: or sip: target, and hang-up ends SIP and WebRTC sessions.

GPT-Live detail. The session.start object rejects unknown fields, model and voice are fixed after start, and the default voice is marin. Azure documents WebSocket and WebRTC but no SIP, so an Azure GPT-Live phone agent needs a media bridge. OpenAI added outbound SIP for GPT-Live on its own API on 2 November 2026. Each WebRTC session bills at least 15 seconds.

Audio formats and phone audio

Sources: Microsoft's Realtime and GPT-Live pages and OpenAI's guide. Push recorded 8 kHz carrier audio through each path in staging.

Function calling and tool behaviour

Realtime. A remote MCP server can be added with a server_url and an approval setting, and the service manages those calls. Out-of-band responses (response.conversation set to none, tagged with response.metadata) run side tasks such as intent classification outside the main conversation. Barge-in relies on response.cancel and conversation.item.truncate, which also deletes the unheard part of the transcript.

GPT-Live. Delegation is either client, where your application runs the work, or responses, where a configured Responses API backend does. Microsoft warns that interrupting speech does not cancel backend work, that a finished backend response does not mean the caller heard it, and that GPT-Live emits no authoritative turn-completed event. Text context can be appended mid-call in chunks of up to 500 tokens.

Either way, test every tool and failure branch, duplicate calls, spoken confirmations against system state, and the end-of-call action.

Latency testing plan

Microsoft's 100 ms and 200 ms figures describe the connection method, not what a caller hears. Measure a gpt-realtime-2.1 baseline now and compare every candidate against it.

Run two or more reasoning-effort settings per candidate; our gpt-realtime-2.1 build guide covers the levels. Keep the old deployment behind a flag and compare recorded calls blind.

Worked cost example: 10,000 four-minute calls a month

Assumptions per call: 2 minutes of caller speech and 1.6 minutes of agent speech at Microsoft's Voice Live estimate of about 10 input and 20 output audio tokens a second for Azure OpenAI models, so 120 x 10 = 1,200 and 96 x 20 = 1,920 tokens; 18,000 cached audio tokens of replayed context across about 12 turns; a 2,000-token prompt sent fresh on the first turn and cached on the 11 later ones (22,000 tokens); 1,000 text output tokens. Prices are Azure Global Standard meters in US East 2. This is arithmetic, not a benchmark.

GPT-Live bills the voice session by time, including silence and backend waits. At $3.00 an hour a four-minute call costs $0.20 for voice. Six delegations to gpt-5.6-luna (18,000 input and 1,800 output tokens at $0.20 and $1.20 per million: $0.0036 + $0.00216) add about $0.0058, for about $0.2058 a call or $2,058 a month. That is close to gpt-realtime-2.1 and well above the mini: a quality choice, not a cost cut, and long holds raise its bill further.

India, data residency and telecom rules

Global deployments may process prompts and responses in any region where the model is deployed; Data Zone deployments stay inside the US, EU or APAC zone. gpt-realtime-2.1 can be deployed as Global Standard in South India, but processing stays global, and its Data Zone rows cover only the US and EU. Voice Live's region table shows gpt-realtime-2.1-regional with regional processing for a Central India resource, though its speech features have separate scopes. Any replacement will likely reach Global Standard first.

Indian SIP trunks reach Realtime SIP in East US 2 or Sweden Central; the Voice Live SIP preview adds Southeast Asia. Record those locations in your DPDP records, limit retention of recordings and transcripts, and re-check when the deployment changes; our DPDP Rules checklist covers the wider duties. Outbound AI calls also fall under TRAI's Third Amendment of 18 September 2026: calls placed by software using artificial voice are application-to-person calls, senders must declare that use and their calling numbers to their operator in advance, and undeclared calls count as unsolicited commercial communication. Our voice agent platform comparison covers the telephony side.

Migration timeline to 25 June 2027

Dates follow the schedule and policy windows. For text models, see our guide to the GPT-5.4 and GPT-5.5 retirement on Azure.

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 handles Realtime, GPT-Live and Voice Live migrations, SIP set-up, recorded-call test sets, latency measurement and retention design. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

1. Model retirement schedule — Microsoft Learn

1. Foundry Models lifecycle and support policy — Microsoft Learn

1. Deprecations — OpenAI

1. Azure OpenAI Service pricing — Microsoft Azure

1. Azure Retail Prices API, gpt-realtime-2.1 Global meters in US East 2 — Microsoft

1. Azure Retail Prices API, GPT-Live, gpt-live-transcribe and Voice Live meters in US East 2 — Microsoft

1. Use the GPT Realtime API via SIP — Microsoft Learn

1. Region availability for Foundry Models sold by Azure — Microsoft Learn

1. What is GPT-Live? — Microsoft Learn

1. Connect Voice Live API to telephony by using SIP — Microsoft Learn

1. API changelog — OpenAI

1. Voice Live API overview — Microsoft Learn

1. Azure OpenAI in Microsoft Foundry Models: working with models — Microsoft Learn

1. Use the GPT Realtime API for speech and audio with Azure OpenAI — Microsoft Learn

1. Realtime conversations guide — OpenAI

1. Realtime API reference — Microsoft Learn

1. Foundry Models sold by Azure — Microsoft Learn

1. GPT-Realtime-2.1 model page — OpenAI

1. Use the GPT Realtime API via WebRTC — Microsoft Learn

1. Use GPT-Live for real-time voice — Microsoft Learn

1. Supported regions for Azure Speech, Voice Live tab — Microsoft Learn

1. TRAI strengthens framework for curbing unsolicited commercial communications, press release of 18 September 2026 — Telecom Regulatory Authority of India