GPT-Live 1.1 now available on AI Gateway
Vercel added OpenAI's full-duplex GPT-Live 1.1 model to AI Gateway, bringing enhanced speech transcription and zero-markup pricing to live voice applications.
By Dillip Chowdary • Nov 19, 2026 • Source: Vercel Blog

At a glance
- Product
- OpenAI GPT-Live
- Version
- 1.1
- Platform
- Vercel AI Gateway
- Platform fee
- Zero markup on provider inference
- Availability
- AI Gateway Live endpoint and playground
- Input mode
- Full-duplex audio and client delegation
In this post
Vercel has added OpenAI's GPT-Live 1.1 to AI Gateway, expanding its hosted model catalog with a full-duplex audio system built for conversational software. According to Vercel Blog's report, the release enables systems to listen continuously while speaking so callers can pause, interrupt, or supply supplemental details mid-response. The release functions directly through existing Live infrastructure and standard client delegation patterns, requiring existing deployments to change only their target model identifier.
This breakdown examines the technical upgrades in transcription and voice consistency, the operational setup on AI Gateway, and the client delegation architecture used for background reasoning tasks. Engineering teams building interactive customer support agents, interactive telephony systems, and voice assistants can evaluate how to route traffic without paying platform inference markups.
01Full-duplex conversational audio with interruption handling
Traditional voice implementations often rely on turn-based audio processing that forces callers to wait until an assistant finishes generating its output. OpenAI built GPT-Live 1.1 with bidirectional audio streams that ingest incoming speech while the model synthesizes outbound audio. When a user interjects or adds clarification mid-reply, the model stops speaking and absorbs the new prompt context immediately.
The underlying communication channel stays active throughout the entire exchange rather than tearing down after individual replies. Developers interact with the runtime via standard WebSocket connections using the Live endpoint on AI Gateway. Because the interface mirrors the previous version, current voice sessions switch over to the updated release without structural architectural rewrites.
02Speech transcription accuracy and voice consistency
The updated release introduces targeted acoustic refinements over the initial GPT-Live deployment, focusing directly on natural language interaction across multilingual environments. Incoming speech transcription demonstrates fewer recognition errors during rapid speech or unexpected caller interruptions. In addition, the vocal output maintains a stable accent, cadence, and tonal identity when alternating across multiple languages.
| Feature | GPT-Live | GPT-Live 1.1 |
|---|---|---|
| Audio Processing | Turn-based or basic duplex | Full-duplex continuous listening |
| Speech Transcription | Standard baseline accuracy | Enhanced precision during interruptions |
| Multilingual Delivery | Variable accent shifts | Consistent accent and voice profile |
| Integration Path | Live endpoint | Same Live endpoint with new model ID |
| Client Delegation | Supported | Supported via delegation-created events |
Voice generation retains its configured personality traits across extended multi-turn conversations without drifting into erratic intonation patterns. This consistency prevents dialect distortion when automated workflows switch vocabularies or handle non-native phrasing from international callers. It also eliminates the need for manual phonetic tuning when serving audiences across different geographical territories.
03Client delegation workflow for secondary models
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
Complex real-time assistants often require secondary language models to execute factual lookups, external reasoning, or database queries while speech synthesis continues uninterrupted. GPT-Live 1.1 enables this separation through client delegation configured inside the initial session instructions. When an inquiry exceeds standard conversational chit-chat, the voice system emits an event requesting specialized external intervention.
- Configure session Enable client delegation within provider options and specify routing conditions in prompt instructions.
- Handle delegation Intercept delegation-created events and invoke a text model such as openai/gpt-6.2-sol through AI Gateway.
- Stream commentary Transmit the generated text back over the commentary channel to let the voice model speak the result.
Host applications retain strict supervisory authority over all delegated tasks, managing individual permissions, user confirmations, and execution cancellations. Background queries to models like openai/gpt-6.2-sol generate their own separate billing entries on AI Gateway. Because the primary voice socket remains connected during background processing, voice session duration charges accrue concurrently.
04Model identifier switch and connection configuration
Deploying GPT-Live 1.1 requires updating the model string inside the standard AI SDK configuration for Vercel AI Gateway. Teams pointing to the base gateway URL initialize the experimental realtime interface with the live API mode. The runtime requests WebSocket configuration tokens directly from the gateway before issuing the session start payload.
```typescript
import { createOpenAI } from '@ai-sdk/openai';Engineers can validate interruption thresholds and test conversational latency directly inside the hosted Vercel model playground before pushing revisions to live environments. Production deployments still require host applications to implement robust event deserialization, raw audio chunking, and session cancellation handlers. These client components ensure that ongoing background tasks terminate promptly whenever a caller disconnects.
05Zero markup pricing and gateway governance
Vercel AI Gateway routes GPT-Live 1.1 requests through an enterprise proxy layer without adding surcharges or per-token markups to upstream OpenAI rates. Teams deploying through Bring Your Own Key credentials also bypass platform inference fees entirely. The gateway logs all incoming audio and delegated text requests under consolidated billing dashboards for unified expense attribution.
The proxy framework includes operational controls such as programmatic routing rules, per-key budget thresholds, and automated retry policies designed to maintain availability during upstream disruptions. Organizations operating under strict compliance standards can activate Zero Data Retention configurations to ensure audio transcripts and generated commentary are never retained in gateway caches.
Note: Production implementations require separate application-level handling for WebSocket streaming, event serialization, and audio shutdown, which are omitted from initial connection snippets. Teams must reference the official integration guides to handle real-time audio buffering and cancellation signals safely.
Frequently asked questions
What is new in OpenAI GPT-Live 1.1 compared to the previous version?
GPT-Live 1.1 delivers more accurate speech transcription and maintains a more consistent accent and voice profile across languages. It also supports full-duplex conversational audio, allowing callers to interrupt, pause, or add details mid-reply.
How do developers upgrade existing GPT-Live applications to GPT-Live 1.1 on AI Gateway?
Existing implementations use the same Live endpoint and client delegation architecture, requiring only an update to the model ID string openai/gpt-live-1.1. WebSocket configurations and event structures remain compatible.
How does client delegation work with GPT-Live 1.1?
When client delegation is enabled in session instructions, the model emits a delegation-created event during complex queries. Applications answer using secondary text models like openai/gpt-6.2-sol and return responses on the commentary channel for the voice model to speak.
Does Vercel charge platform fees for running GPT-Live 1.1 on AI Gateway?
No, Vercel AI Gateway passes through OpenAI provider pricing without markups and charges zero platform fees on inference, including on Bring Your Own Key requests. Delegated model calls are billed separately from ongoing voice sessions.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement