# gpt-realtime-2

gpt-realtime-2 lets voice agents think before they speak. Configurable reasoning effort matches compute to the conversation, so simple turns stay fast and hard problems get real reasoning.

- **Model ID:** `openai/gpt-realtime-2`
- **Type:** realtime
- **Providers:** openai
- **Context window:** Not applicable
- **Maximum output tokens:** 0
- **Pricing:** $4/1M input tokens, $24/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/gpt-realtime-2

## 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',
  });

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

## About

gpt-realtime-2 launched on May 7, 2026 as the successor to `gpt-realtime-1.5` and OpenAI's most capable realtime voice model. The design goal: voice agents that reason through problems as conversations unfold, instead of only routing or reciting.

Reasoning effort is configurable across levels from minimal to xhigh, so you choose how much thinking each session gets. Low effort is the default and keeps responses quick. Higher effort raises answer quality on complex workflows, at the cost of latency and output tokens. When a caller interrupts mid-thought, gpt-realtime-2 discards the current chain of thought and starts a fresh turn.

Conversation mechanics improved alongside reasoning. Preambles let an agent acknowledge the request out loud while tools run, parallel tool calls fire multiple requests at once, and recovery behavior catches tool failures rather than freezing the call. OpenAI reports higher scores than GPT-Realtime-1.5 on audio reasoning and instruction-following benchmarks.

## What to consider

Reasoning effort is the main tuning decision. Low is the default and keeps latency down for most production voice agents; higher settings improve reasoning at the cost of latency and output token usage. Realtime support on AI Gateway is in beta, with a WebSocket integration through the AI SDK's realtime hook and short-lived server-minted tokens.

## When to use

### Best For

- Voice agents that solve multi-step problems during the call
- Complex support workflows with tool chains, lookups, and decisions mid-conversation
- Mixed-difficulty sessions, using reasoning effort to match compute to each turn
- Agents that keep talking naturally while tools execute in the background

### Consider Alternatives When

- Most turns are simple and latency critical; `gpt-realtime-1.5` remains the flagship for standard voice agents
- Voice features are high volume and cost sensitive; `gpt-realtime-mini` handles lightweight interactions
- You only need transcription or speech generation; dedicated models like `gpt-4o-transcribe` and `tts-1` fit better

## Best for

- **Multi-step voice problem solving:** Agents that reason through hard requests as the conversation unfolds
- **Complex tool workflows:** Parallel tool calls, lookups, and decisions handled mid-conversation
- **Mixed-difficulty sessions:** Reasoning effort settings match compute and latency to each turn
- **Natural conversation under load:** Preambles and recovery behavior keep calls flowing while tools run

## Consider alternatives

- **Latency-critical simple turns:** `gpt-realtime-1.5` keeps standard voice agents fast without reasoning overhead
- **High-volume voice features:** `gpt-realtime-mini` delivers realtime interaction at a cost built for scale
- **Transcription or speech only:** `gpt-4o-transcribe` and `tts-1` cover single-direction audio jobs

## Frequently asked questions

### What reasoning effort levels does gpt-realtime-2 support?

Levels range from minimal to xhigh, with low as the default. Higher effort improves reasoning quality on complex turns and increases latency and output token usage, so start low for most production voice agents.

### How do I access gpt-realtime-2 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. Your AI Gateway API key never reaches the client.

### What happens when a caller interrupts gpt-realtime-2 mid-response?

gpt-realtime-2 discards the current chain of thought and starts a new turn. Interruption handling is part of the conversation mechanics, alongside preambles and recovery from failed tool calls.

### How does gpt-realtime-2 differ from gpt-realtime-1.5?

gpt-realtime-2 is the more capable successor, adding configurable reasoning effort, stronger instruction following, and more reliable tool use. GPT-Realtime-1.5 remains a strong flagship option for standard voice agents where latency is the priority.

### How is gpt-realtime-2 priced?

Pricing is listed on this page. Rates reflect the providers routing through AI Gateway and update when providers change list prices.

### Does gpt-realtime-2 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 responses.

## Links

- [Model page](https://vercel.com/ai-gateway/models/gpt-realtime-2)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://developers.openai.com/api/docs/models/gpt-realtime-2)