Realtime Voice APIs Compared: OpenAI, Gemini, xAI, AWS
Compare realtime voice APIs as of October 2026: OpenAI Realtime and GPT-Live, Gemini 3.8 Live, Grok Voice 3.0, Nova 2 Sonic and Azure. Prices and limits.
A realtime voice API gives a developer a live, two-way audio session with a speech model, usually over WebSocket or WebRTC. As of October 2026, the main model-level options are OpenAI's Realtime API and GPT-Live, Google's Gemini 3.8 Live, xAI's Grok Voice 3.0 and Amazon's Nova 2 Sonic, with Microsoft's Azure Voice Live as an interface. Platforms such as ElevenLabs Agents and Vapi add orchestration on top, and their prices are not comparable with model prices.
Models and platforms are different products
A model-level API gives you the speech model and nothing else. You handle the telephony, the tool layer, the turn logic and the business rules. A platform sells those parts together, usually for a per-minute fee on top of the model cost. Both can be the right choice. The question is which parts you want to own.
Realtime voice APIs at a glance
The table lists the model-level options and the main platform. Prices are as published on each provider's page in October 2026, and the units differ, so read each row on its own terms.
| Provider and product | Transport and phone | Published price | Stated limits |
|---|---|---|---|
| OpenAI gpt-realtime-2.1 | WebRTC and WebSocket; SIP through a trunking provider | Audio $32 input and $64 output per 1M tokens | 128,000-token context; 32,000 max output tokens |
| OpenAI GPT-Live-1 | Live endpoint only (v1/live/sessions); WebRTC recommended; WebSocket, telephony and SIP | $0.05 per minute, billed per second | Structured outputs and fine-tuning not supported |
| Google Gemini 3.8 Live | Live API over WebSocket (WSS); WebRTC is not described for the API itself | Audio $0.005 per minute in, $0.018 per minute out; free tier | 131,072 input and 65,536 output tokens |
| xAI Grok Voice 3.0 | WebSocket | $0.08 per minute, plus $0.004 per text input | Sessions up to 120 minutes |
| Amazon Nova 2 Sonic | Bidirectional streaming | Not verified; see below | 8-minute connection limit, renewed with a session-continuation pattern |
| Microsoft Azure Voice Live | WebSocket, compatible with Azure OpenAI Realtime events | Three price tiers set by the model | 140-plus locales for speech to text |
| ElevenLabs Agents | Platform; SIP trunking and a native Twilio integration | $0.08 per minute beyond included minutes; LLM billed separately | Included minutes and concurrency vary by plan |
OpenAI: Realtime API and GPT-Live
OpenAI runs two product lines, and they are easy to confuse. The Realtime API uses the gpt-realtime-* models at /v1/realtime. It went generally available on August 28, 2025. gpt-realtime-2 arrived on May 7, 2026 with configurable reasoning, and gpt-realtime-2.1 and gpt-realtime-2.1-mini followed on July 6, 2026. The mini model is a faster, lower-cost distilled reasoning model.
GPT-Live is separate. gpt-live-1 is served only through /v1/live/sessions, is billed per minute and became generally available in the API on September 10, 2026. On October 16, 2026, OpenAI added nine voices for German, French, Hindi, Japanese and Korean, along with prompt-to-voice custom voices for GPT-Live. Creating a voice is free, and speech is billed as normal session time.
For phone calls, OpenAI connects through a SIP trunk. The trunk points at sip:$PROJECT_ID@sip.api.openai.com;transport=tls, or at sip-eu.api.openai.com for EU residency. OpenAI sends a realtime.call.incoming webhook, and endpoints let you accept, reject, transfer or hang up the call. Regional processing endpoints carry a 10 percent price uplift for models released on or after March 5, 2026. The OpenAI pricing page has the token rates for each model, and the SIP guide covers the call flow.
Google: Gemini 3.8 Live
Gemini 3.8 Live (gemini-3.8-live) is marked Stable on its model page, which was last updated on September 15, 2026. Google describes it as "the default option for most low-latency voice agent experiences." The model accepts text, images, audio and video, and returns text and audio. It supports function calling, search grounding and thinking, but not caching or the Batch API.
The Live API is stateful over WebSocket. It takes 16-bit PCM audio at 16 kHz and returns audio at 24 kHz. Users can interrupt the model at any time, and the API returns transcripts of input and output. It lists 70 supported languages. The pricing page includes a free tier for these models. The Live API page states no session limits. Check the current terms before you build on the free tier.
xAI: Grok Voice 3.0
xAI released Grok Voice 3.0 on October 14, 2026. Its Voice Agent API costs $0.08 per minute, or $4.80 an hour, the same rate as Grok Voice 2.0. xAI's pricing page marks 2.0 as deprecated. The API docs describe a WebSocket connection with a 120-minute maximum session, in the us-east-1 region. Sessions that use server voice activity detection are billed for session duration, and push-to-talk sessions are billed for the audio sent and received.
xAI's own post says the model scores 86.9 on the Artificial Analysis index and 71.6 percent on tau-Voice. Those are vendor-cited figures. The docs checked for this article do not mention WebRTC or SIP, so confirm telephony support before you plan a phone deployment.
Amazon and Microsoft: Nova 2 Sonic and Azure Voice Live
Nova 2 Sonic is a bidirectional streaming model. It supports seven languages: English in US, UK, Indian and Australian voices, plus French, Italian, German, Spanish, Portuguese and Hindi. AWS documents intelligent turn-taking, graceful interruption handling and asynchronous tool handling, which lets the assistant keep speaking while a tool runs. Each connection is limited to 8 minutes, and the documented pattern is to renew the session. The AWS pages checked for this article do not list a Sonic price, and third-party figures conflict, so this article gives none.
Azure Voice Live is an interface rather than a model of its own. It is WebSocket-based and compatible with Azure OpenAI Realtime events. Its model menu includes gpt-realtime-2.1, gpt-realtime-2.1-mini and gpt-realtime-1.5, along with GPT-5 family models and gpt-4.1, which gain audio through Azure speech. Microsoft lists 140-plus locales for speech to text and 600-plus standard voices across 150-plus locales. Custom voices and avatars are limited and billed separately.
Platforms on top
Platforms bundle orchestration, telephony and turn-taking, and they bill for it. The published rates below are for October 2026.
- Deepgram Voice Agent API. Standard is $0.075 per minute. It is $0.065 with your own speech synthesis, $0.050 with your own language model and synthesis, and $0.163 on Advanced. Billing runs on WebSocket connection time.
- ElevenLabs Agents. $0.08 per minute beyond included minutes, with a burst rate of $0.16 per minute. The included minutes run from 15 on Free to 12,375 on Business. ElevenLabs announced the cut from $0.10 on May 7, 2026. Its LLM usage is billed separately.
- Ultravox. $0.05 per minute after the first 30 free minutes. The Pro plan is $100 a month with no hard concurrency cap, and the open-weights project is MIT licensed.
- Vapi. $0.05 per minute for hosting, with speech and model costs passed through at provider cost. The Core plan is $29 a month, and its HIPAA add-on is $2,000 a month.
- Retell AI. $0.07 to $0.31 per minute, with the voice engine at $0.055 per minute and telephony at $0.015 per minute.
- LiveKit Cloud. The Build plan is free and includes 1,000 agent-session minutes and 5 concurrent sessions. Agent sessions cost $0.0100 per minute, plus inference at model rates.
What the leaderboard says
Artificial Analysis scores speech-to-speech models on an index and reports first-audio time. The rows below are a selection from its October 2026 table.
| Model | Index score | TTFA (seconds) | Benchmark cost per hour (USD) | |---|---|---|---| | Grok Voice 3.0 | 86.9 | 0.81 | 4.80 | | GPT-Live-1 (Astra label, medium) | 83.2 | 1.34 | 5.83 | | Gemini 3.8 Live | 76.0 | 1.18 | 0.84 | | GPT-Realtime-2.1 (High) | 75.2 | 1.21 | 10.75 |
The page's description of its index weighting contradicts itself, so do not read the score as a weighted average of anything in particular. The page also scores no cascaded systems, so it cannot rank native models against pipelines. The speech-to-speech comparison covers that question, and the latency guide covers what TTFA does and does not measure.
How to choose
- Phone or browser? Check SIP and WebRTC support before anything else. The docs for each model differ.
- Which languages? Use the vendor's own list. Counts range from 7 for Nova 2 Sonic to 70 for Gemini's Live API, and some vendors state none.
- Session limits. xAI allows 120-minute sessions, while Nova 2 Sonic connections run 8 minutes before renewal.
- Cost unit. Per token, per minute, per connection time or bundled. Convert each to your expected call length.
- Tools and reasoning. Confirm function calling and any reasoning setting for the exact model.
- Data residency. OpenAI offers an EU SIP endpoint, and regional processing carries a 10 percent uplift.
- Build or buy. A platform adds orchestration for a per-minute fee. A model API leaves that work to you.
Where recorded voices fit
OndaLabs has no realtime or voice-agent API, and no telephony. It is a studio for generated audio. Its fit with these APIs is the recorded layer around them: greetings, hold audio, disclosure lines and a consistent brand voice. The studio has 30 prebuilt voices that work in every supported language. Custom voices, designed or cloned for 1,500 credits each, are available on the Creator plan (3 voices), Pro (10) and Business (30). The text to speech studio is where those prompts get made.
The Free plan includes 2,000 credits a month, about two minutes of audio. Commercial use is included from Starter, which costs $9 a month for 170,000 credits. Top-ups are 100,000 credits for $5 and 500,000 credits for $24. The pricing page has the full plan list.
Frequently asked questions
Which realtime voice API is cheapest per minute?
It depends on what the price covers. GPT-Live-1 is $0.05 per minute with backend usage billed separately, and xAI's Grok Voice 3.0 is $0.08 per minute. Platforms add their own per-minute fees, and token-priced models cost differently again. Estimate each option for your real call length and context size before comparing.
Does GPT-Live replace the Realtime API?
No. GPT-Live is a separate product line with its own endpoint, /v1/live/sessions, and per-minute billing. The Realtime API continues to use the gpt-realtime-* models at /v1/realtime. Pick the one that matches your architecture and billing model.
Which realtime voice APIs support SIP phone calls?
OpenAI supports SIP for both the Realtime API and GPT-Live, through a trunking provider such as Twilio. ElevenLabs Agents offers SIP trunking and a Twilio integration, and LiveKit agents support SIP. The xAI docs checked for this article do not mention SIP, so confirm it directly with the provider.
Is there a free realtime voice API?
Google's pricing page lists a free tier for Gemini 3.8 Live and related models. Ultravox includes 30 free minutes before its per-minute rate applies, and LiveKit's Build plan includes 1,000 agent-session minutes. Free tiers have limits and change over time, so check the current terms before you build.
How many languages do realtime voice APIs support?
It varies widely. Gemini's Live API lists 70 supported languages and Azure lists 140-plus locales for speech to text. Nova 2 Sonic supports seven, and OpenAI's pages checked for this article give no count. Confirm the languages you need against each vendor's own documentation.