GPT-Live 1 now available on AI Gateway: GPT-Live 1 from OpenAI is now

Vercel's AI Gateway now supports GPT-Live 1, OpenAI's full-duplex voice model, expanding the routing and observability layer to cover real-time audio workloads alongside the text and image models it already handles. The addition means developers can route voice traffic through a single control plane rather than managing a separate integration path for conversational audio.

This piece covers what GPT-Live 1 brings to AI Gateway, how full-duplex audio differs from turn-based voice, and what the client-delegation feature means for teams building layered voice applications. It is written for developers and platform engineers who manage model routing, cost, and reliability on Vercel's infrastructure.

What shipped in GPT-Live 1 now available on AI Gateway

Vercel added GPT-Live 1 to AI Gateway, making OpenAI's full-duplex voice model accessible through the same proxy and observability tooling that already covers text and multimodal requests. Full duplex means GPT-Live 1 can both listen and produce speech simultaneously over a single persistent connection, rather than waiting for the user to stop talking before it begins processing.

Most voice models in production today rely on turn detection: a silence threshold or endpoint signal tells the model the user has finished, and only then does the model begin generating a response. GPT-Live 1 removes that hard boundary, so the audio channel remains bidirectional throughout a conversation. The practical effect is that users can pause mid-sentence, introduce a correction, or add context while the model is still speaking, without triggering an awkward restart.

What improved in GPT-Live 1 now available on AI Gateway

The headline capability difference between GPT-Live 1 and conventional voice pipelines is the elimination of turn-based latency introduced by endpoint detection. Traditional turn-detection models must accumulate silence, classify it as an intentional pause, then fire a generation request — a pipeline that introduces both latency and errors when users speak with natural hesitation. GPT-Live 1 processes input and output concurrently.

The source material does not provide benchmark numbers, price-per-minute figures, or latency measurements to compare against prior versions, so no numeric table can be constructed from available facts. Developers evaluating GPT-Live 1 against their existing voice stack should instrument their own latency and interruption-handling metrics through AI Gateway's logging layer once traffic is live to establish a baseline for their specific use case.

What you gain from GPT-Live 1 now available on AI Gateway

Client delegation is the second major capability GPT-Live 1 introduces. It lets developers select the background reasoning or processing model independently from GPT-Live 1 itself, decoupling the voice interface layer from the underlying model doing the heavy lifting. Teams can therefore swap or upgrade the background model without changing the audio integration.

For platform teams using AI Gateway, the combination of full-duplex audio and client delegation opens a cleaner separation of concerns in voice architecture: GPT-Live 1 handles the real-time audio channel, and a separately configured model handles tasks like retrieval, tool calls, or structured reasoning. That split also means cost and performance can be tuned at each layer independently, rather than accepting a single model's trade-offs for both the conversational surface and the compute-intensive work behind it.

How to get GPT-Live 1 now available on AI Gateway

GPT-Live 1 is available through Vercel's AI Gateway now. Developers already using AI Gateway can add it by updating to the latest version of the Vercel AI SDK and pointing their voice client at the gateway endpoint with the GPT-Live 1 model identifier. The commands below reflect the SDK package update path; consult the Vercel AI Gateway dashboard to confirm the exact model string in your account's model list.

```bash
# Update the Vercel AI SDK to the latest version
npm install ai@latest

Or with pnpm

// Point your voice client at GPT-Live 1 through AI Gateway
const model = 'gpt-live-1'; // confirm exact identifier in your AI Gateway dashboard

To set GPT-Live 1 as the default voice model for a project, update the model field in your AI Gateway configuration file or environment variable rather than hardcoding the identifier in each request. Client delegation is configured separately: specify the background model in the delegation parameter when initializing the session.

What to watch after GPT-Live 1 now available on AI Gateway

The client-delegation feature deserves close attention as OpenAI and Vercel publish more detail on supported background models and any constraints on which model combinations are permitted. Developers building production voice agents will want to know whether delegation supports tool-calling models, whether there are latency implications when the background model is on a different region, and how AI Gateway handles observability across the two-model split.

Interruption handling at scale is the other area to monitor. Full-duplex audio that supports mid-speech user input adds complexity to conversation state management: systems must decide when an interruption should halt generation, when it should be folded into the ongoing response, and how that state is surfaced through AI Gateway's logging. Teams shipping voice features should test interruption edge cases early rather than discovering state management gaps in production.