Best Real-Time Speech-to-Speech AI APIs for Phone Calls (2026)

The best real-time speech-to-speech AI services for phone calls in 2026 are Retell AI, the OpenAI Realtime API, OpenAI GPT-Live-1, the Google Gemini Live API, Amazon Nova 2.5 Sonic, the Deepgram Voice Agent API and the AssemblyAI Voice Agent API.

Retell is the pick when the calls are production phone calls, because it wraps a speech-to-speech model in the telephony, transfers and testing a phone line needs. The other six are model or pipeline APIs. They give you the voice loop and leave most of the phone system, the human handoff and the QA to you.

Retell AI publishes this list and is one of the options compared. Competitor prices and features come from each vendor's public pricing, model or documentation page, read on October 22, 2026, and linked where they are cited.

How we picked and compared the options. We included every API that accepts live audio and returns speech in real time and that a team could put behind a phone number: native speech-to-speech models from OpenAI, Google and Amazon, two bundled voice agent APIs from Deepgram and AssemblyAI, and Retell as the platform layer. We compared them on five criteria (below) using only vendor-published information. We did not place test calls or run a latency test, so the article gives no scores and no latency figures.

TL;DR

• Native speech-to-speech models: OpenAI gpt-realtime-2.1, OpenAI GPT-Live-1, Gemini 3.8 Live and Amazon Nova 2.5 Sonic take audio in and return audio out in one model.
• Bundled voice agent APIs: Deepgram and AssemblyAI run speech-to-text, an LLM and text-to-speech behind one WebSocket and one per-minute rate.
• Lowest published per-minute audio figures: Gemini 3.8 Live, at $0.005 a minute in and $0.018 a minute out, which Google converts from $3 and $12 per 1M audio tokens.
• Flat per-minute rates: GPT-Live-1 at $0.05 a minute for the voice layer (backend model billed separately), Deepgram from $0.075 a minute and AssemblyAI at $0.075 a minute all in.
• Direct SIP: OpenAI accepts SIP calls on both its Realtime API and its Live sessions, and AssemblyAI takes SIP through your own Twilio account. For Gemini Live, Google's docs point to partner integrations such as LiveKit and Pipecat.
• Production phone agents: Retell runs OpenAI's realtime models, Grok Voice or a standard LLM pipeline, and adds warm transfer, IVR navigation, simulation testing and post-call analysis.

How do the 7 APIs compare?

The split that matters is whether you are buying a model or a phone agent. The table shows what each option is, how a phone call reaches it, and its published price on October 22, 2026.

How we compared each API

We looked at five things a phone call exposes and a browser demo hides:

1. Telephony path. Can a PSTN call reach the model without you running a media server?
2. Turn-taking under real conditions. Does the vendor document interruption handling for live calls?
3. Tool calls during speech. Can the agent look up an order while it keeps talking?
4. Handoff to a human. Is warm transfer built in, or is it your code?
5. Cost you can forecast. Is billing per minute or per token, and what is outside the meter?

Vendor latency claims are measured differently and are not comparable, so none appear here.

Retell AI

Retell AI is the best choice when a speech-to-speech model has to survive real phone calls, because it handles the call around the model: numbers, handoff, voicemail and testing.

Retell's API lists speech-to-speech options including OpenAI's gpt-realtime family (gpt-realtime-2.1, gpt-realtime-2 and others) and Grok Voice 3.0, as one choice among its supported models. You can also run an agent on a standard pipeline with the LLM of your choice, and keep the same telephony, transfers and testing either way. See the model options in the API reference.

For evidence on hard calls, Cekura ran a 414-call benchmark of six voice platforms (Retell, ElevenLabs, Pipecat, LiveKit, Synthflow and Vapi) on one regulated Medicare workflow. Retell passed 22 of 23 evaluators and led with 95.7% workflow accuracy, counting only checks that passed on all three runs. The hard cases included a caller who interrupts the disclosure and demands plan details, a family member calling on behalf of a beneficiary, and a caller asking the agent to promise that a plan covers their medication. It is a platform benchmark on one workflow, not a comparison of speech-to-speech models.

What you get on top of the model:

• Telephony: buy US or Canada numbers in the dashboard, or import your own over SIP trunking, including Twilio and Vonage.
• Handoff: cold, warm and agentic call transfer to a person, with a contextual briefing on warm transfers.
• Outbound telephony: voicemail detection and IVR navigation through other companies' phone menus.
• Testing and QA: simulation testing before launch, then post-call analysis and AI quality assurance on production calls.

Pricing: from $0.07 per minute on Retell's pricing page.

Watch out for: Retell is a platform, not a raw model endpoint. If you want to own every frame of the audio pipeline yourself, start with a model API instead.

OpenAI Realtime API (gpt-realtime-2.1)

The OpenAI Realtime API is the pick for teams that want a raw speech-to-speech model with GPT reasoning and will build the rest of the phone stack themselves.

gpt-realtime-2.1 takes audio in and returns audio out, supports configurable reasoning effort and function calling, and updates gpt-realtime-2 with better alphanumeric recognition, silence and noise handling, and interruption behavior. OpenAI lists the same prices for both models.

It also answers a phone call without a separate bridge. Per OpenAI's SIP guide, you point a SIP trunk at OpenAI's SIP endpoint, accept or reject each call with its call ID, and can transfer it with a SIP REFER.

Pricing: $32 per 1M audio input tokens and $64 per 1M audio output tokens, plus $4 and $24 per 1M text tokens. Higher reasoning effort can increase latency and output token usage.

Watch out for: token billing grows with conversation length because context accumulates. Recording, QA, retries and human escalation logic are yours to build.

OpenAI GPT-Live-1

GPT-Live-1 is the best fit when natural turn-taking matters more than having reasoning inside the voice model. It is a full-duplex model that listens and speaks at the same time, and it delegates reasoning and tool use to a backend agent.

It runs on OpenAI's Live sessions endpoint, not the Realtime endpoint, so it is not a drop-in swap for gpt-realtime-2.1. Per OpenAI's voice and SIP guide, inbound calls arrive over SIP and outbound SIP calling is available through the Live API once it is enabled for your organization. Ringing is limited to 3 minutes and a connected call to 2 hours.

Pricing: $0.05 per minute of session, billed per second, with backend model and tool usage billed separately.

Watch out for: concurrency is capped by tier, at 50 sessions on Build, 300 on Launch and 500 on Grow, and inbound SIP needs GPT-Live SIP support enabled for your project.

Google Gemini Live API

The Gemini Live API has the lowest per-minute audio figures on published prices. Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking were generally available on September 15, 2026, and both list $0.005 a minute for audio in and $0.018 a minute for audio out on Google's pricing page. Those per-minute figures sit beside $3 and $12 per 1M audio tokens, so they are a conversion, not a separate meter.

The Extended Thinking model supports background reasoning during the call, and Google Search grounding is supported on the paid tier, with 5,000 free search requests a month and then $14 per 1,000.

Watch out for: Google's Live API guide lists partner integrations such as LiveKit and Pipecat and describes no SIP trunk path, so plan on a bridge in front of the Live API. Output pricing includes thinking tokens, so the per-minute figure is a floor.

Amazon Nova 2.5 Sonic

Amazon Nova 2.5 Sonic is the default choice for teams already building on AWS Bedrock. AWS made it generally available on October 5, 2026, with improved reasoning, instruction following and tool-calling accuracy, lower latency, expressive voices in seven languages, controllable turn-taking, asynchronous tool calling and a 256K context window.

AWS's Strands Bidi Agents framework, now generally available, is its route to building a voice agent in a few lines of code.

Pricing: AWS lists Nova 2.5 Sonic at the same price as Nova 2 Sonic. Check the Amazon Nova pricing page for current per-token rates, which load dynamically.

Watch out for: it is available in four Bedrock regions (N. Virginia, Oregon, Stockholm, Tokyo), and seven languages is narrower than some other options here. The announcement does not describe a phone path, so confirm how calls reach the model before you commit.

Deepgram Voice Agent API

Deepgram's Voice Agent API is the pick for teams that want Deepgram's speech recognition with the option to bring their own LLM or voice. It runs speech-to-text, an LLM and text-to-speech behind one WebSocket and one per-minute meter, and Deepgram's docs include inbound and outbound telephony agent guides for Twilio and Plivo.

Pricing: $0.075 a minute on the Standard pay-as-you-go tier and $0.163 on Advanced, with lower rates if you bring your own LLM or TTS.

Watch out for: it is a cascaded pipeline, not a native speech-to-speech model. Pay-as-you-go accounts are limited to 45 concurrent Voice Agent connections (60 on Growth), which matters the first time a campaign dials out.

AssemblyAI Voice Agent API

AssemblyAI's Voice Agent API is the pick when transcription accuracy decides whether the agent works, for example on names, account numbers and addresses. It runs AssemblyAI's Universal-3.6 Pro Realtime speech model, an LLM and TTS behind one WebSocket.

Pricing: a flat $4.50 an hour ($0.075 a minute) covering speech-to-text, the LLM and TTS, billed per session minute, per AssemblyAI's pricing reference. For phone calls, AssemblyAI's docs say Twilio passes calls to AssemblyAI over SIP, so there is no media server or webhook to run. The SIP trunk and number sit in your own Twilio account.

Watch out for: like Deepgram it is cascaded, not native audio-to-audio, so you get a transcript at each turn but not a model that hears tone directly.

Should you use a native speech-to-speech model or a cascaded pipeline?

Use a native speech-to-speech model when tone and fast turn-taking matter more than control over each step, and a cascaded pipeline when you need to see and correct what the agent heard. A native model takes audio in and returns audio out, with no separate transcription step. A cascaded pipeline gives you a transcript at every turn, which makes it easier to debug a wrong account number or enforce a compliance script.

Either way, a phone line adds work the model does not do: interruption handling, transfer recovery and testing against calls that go off script. For how the two designs differ in more detail, read how real-time voice AI works. If you are picking the reasoning model behind a cascaded agent, see the best LLMs for voice AI agents and the best speech-to-text models.

How do you pick the right API for your calls?

Start from where the call lands and who owns the failure. Four questions settle it:

1. Do the calls come from the phone network? If yes, rule out anything that needs a bridge you do not want to run, or pick a platform that includes telephony. Our guide to SIP trunking and modern telephony covers the carrier side.
2. What happens when the agent does not know? If a person must take over with context, you need warm transfer, not just a hang-up endpoint.
3. Can you forecast the bill? Per-minute pricing (GPT-Live-1, Deepgram, AssemblyAI) is easier to budget than per-token pricing, where long calls cost more per minute.
4. Who fixes a failed call? If your operations team, not your engineers, will tune prompts and flows, choose a platform with testing and versioning built in.

FAQ

What is a speech-to-speech AI API?

A speech-to-speech AI API takes a caller's audio and returns spoken audio from one model, without a separate transcription and voice step. OpenAI gpt-realtime-2.1, GPT-Live-1, Gemini 3.8 Live and Amazon Nova 2.5 Sonic work this way.

Which speech-to-speech API is cheapest for phone calls?

Gemini 3.8 Live lists the lowest per-minute audio figures, at $0.005 a minute for audio in and $0.018 a minute for audio out, converted by Google from $3 and $12 per 1M audio tokens. Your real cost adds telephony, any bridge you run, and thinking tokens, which bill as output. AWS lists Nova 2.5 Sonic at the same price as Nova 2 Sonic, and OpenAI bills the Realtime API per token, so compare on your own call length.

Can I connect the OpenAI Realtime API to a phone number?

Yes. Point a SIP trunk from a carrier such as Twilio at OpenAI's SIP endpoint, then accept each call with its call ID. Transfers use a SIP REFER.

Does Retell support speech-to-speech models?

Yes. Retell's API lists speech-to-speech options including OpenAI's gpt-realtime family and Grok Voice 3.0, and a Retell agent can instead run on a standard pipeline with your chosen LLM. The telephony and transfer features are the same either way.

What happens when the AI agent cannot answer?

On a model API, you write the escalation logic. On Retell, the agent can do a cold, warm or agentic call transfer to a person, and a warm transfer passes a contextual briefing so the caller does not repeat themselves.

Can I keep my existing phone number?

On Retell, yes: you can import an existing number over SIP trunking instead of buying a new one. On the model APIs, it depends on the SIP trunk or bridge you run.

Try it on your own calls

A demo script will not tell these APIs apart; your own calls will. Build a Retell agent free, pick a speech-to-speech or standard model, and run it against the calls your team finds hardest.