# gpt-realtime-2.1

gpt-realtime-2.1 sharpens the parts of a voice call that break first: spoken codes and account numbers, background noise and silence, and callers who interrupt mid-response.

- **Model ID:** `openai/gpt-realtime-2.1`
- **Type:** realtime
- **Providers:** openai
- **Context window:** 128,000
- **Maximum output tokens:** 32,000
- **Pricing:** $4/1M input tokens, $24/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/gpt-realtime-2.1

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { gateway } from '@ai-sdk/gateway';

export async function POST() {
  const { token, url } = await gateway.experimental_realtime.getToken({
    model: 'openai/gpt-realtime-2.1',
  });

  return Response.json({ token, url, tools: [] });
}
```

## About

gpt-realtime-2.1 became available on July 9, 2026 as an update to `gpt-realtime-2`. OpenAI improved three behaviors that decide whether a voice call succeeds: alphanumeric recognition, silence and noise handling, and interruption handling. Speech-to-speech interaction, configurable reasoning effort, instruction following, and tool use all carry over.

Alphanumeric recognition covers order numbers, account IDs, postal codes, and confirmation codes, which is where voice agents most often lose a call. Better silence and noise handling reduces false turn boundaries in loud rooms and over phone lines, so the agent talks over the caller less. Improved interruption behavior lets a caller cut in without derailing the session.

Reasoning effort stays configurable, so you decide how much thinking each session gets. Higher effort raises answer quality on complex turns and increases latency and output token usage. OpenAI also improved caching across the realtime voice models, which brings tail latency down. See the live metrics on this page for current numbers.

gpt-realtime-2.1 takes text, audio, and image input and returns text and audio, with a context window of 128K tokens and up to 32K tokens of output. Function calling works mid-conversation, so an agent can look up an order or update a record while the caller is still talking. Through AI Gateway, you get the same authentication, observability, and spend controls as your text models, with no markup on provider pricing.

## What to consider

Reasoning effort is the main tuning decision. Higher effort improves answers on complex turns and raises latency and output token usage, so start low for production voice agents and increase it only where your transcripts show the model struggling.

Realtime support on AI Gateway is in beta through AI SDK 7. The integration runs over WebSocket: your server mints a short-lived token, and the browser connects with the AI SDK's realtime hook, which handles microphone capture and playback. Your AI Gateway API key never reaches the client.

## When to use

### Best For

- **Production Voice Agents:** Support and phone workflows that need reliable instruction following and tool use
- **Spoken Identifier Capture:** Order numbers, account codes, and postal codes read aloud by callers
- **Noisy Call Environments:** Contact centers, vehicles, and public spaces where background sound is constant
- **Interruption-Heavy Conversations:** Callers who cut in mid-response and expect the agent to keep up
- **Mixed-Difficulty Sessions:** Reasoning effort tuned to match compute to the task in front of the agent

### Consider Alternatives When

- **Live Transcription Only:** `gpt-realtime-whisper` streams text from audio without a spoken reply
- **High-Volume Voice Features:** `gpt-realtime-mini` covers lightweight interactions at a lower rate
- **Recorded Audio Jobs:** `gpt-4o-transcribe` and `whisper-1` transcribe files rather than live streams
- **Speech Generation Only:** `tts-1` and `tts-1-hd` turn written text into spoken audio

## Best for

- **Production Voice Agents:** Support and phone workflows that need reliable instruction following and tool use
- **Spoken Identifier Capture:** Order numbers, account codes, and postal codes read aloud by callers
- **Noisy Call Environments:** Contact centers, vehicles, and public spaces where background sound is constant
- **Interruption-Heavy Conversations:** Callers who cut in mid-response and expect the agent to keep up
- **Mixed-Difficulty Sessions:** Reasoning effort tuned to match compute to the task in front of the agent

## Consider alternatives

- **Live Transcription Only:** `gpt-realtime-whisper` streams text from audio without a spoken reply
- **High-Volume Voice Features:** `gpt-realtime-mini` covers lightweight interactions at a lower rate
- **Recorded Audio Jobs:** `gpt-4o-transcribe` and `whisper-1` transcribe files rather than live streams
- **Speech Generation Only:** `tts-1` and `tts-1-hd` turn written text into spoken audio

## Frequently asked questions

### What changed between gpt-realtime-2.1 and GPT-Realtime-2?

Alphanumeric recognition, silence and noise handling, and interruption behavior all improved. Speech-to-speech interaction, configurable reasoning effort, instruction following, and tool use carry over from GPT-Realtime-2.

### How do I access gpt-realtime-2.1 through AI Gateway?

Realtime support is in beta through AI SDK 7. Mint a short-lived token on your server with the AI SDK's gateway provider, then connect from the browser with the AI SDK's realtime hook over WebSocket. The hook manages microphone capture and audio playback.

### How should I set reasoning effort for gpt-realtime-2.1?

Start low for production voice agents. Higher effort improves reasoning on complex turns and increases both latency and output token usage, so raise it only for the sessions or turns where transcripts show the agent struggling.

### Why does alphanumeric recognition matter for a voice agent?

Order numbers, account IDs, postal codes, and confirmation codes are the details a call turns on. gpt-realtime-2.1 captures them more reliably than GPT-Realtime-2, which reduces the repeat-back loops that frustrate callers.

### Can gpt-realtime-2.1 call tools during a live conversation?

Yes. Function calling works mid-conversation, so the agent can look up an order or update a record while the caller is still talking, then answer with the result.

### What inputs and outputs does gpt-realtime-2.1 support?

Text, audio, and image input, with text and audio output. The context window is 128K tokens, with up to 32K tokens of output.

### How is gpt-realtime-2.1 priced?

Pricing is listed on this page and covers text, audio, and image tokens separately. Rates reflect the providers routing through AI Gateway and update when those providers change list prices.

### Does gpt-realtime-2.1 support zero data retention through AI Gateway?

Zero Data Retention is not currently available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

### What are typical latency characteristics?

This page shows live performance metrics measured across real AI Gateway traffic. Higher reasoning effort settings add thinking time before the agent speaks.

## Links

- [Model page](https://vercel.com/ai-gateway/models/gpt-realtime-2.1)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://developers.openai.com/api/docs/models/gpt-realtime-2.1)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.