# Building with GPT-Live and Image Generation | DevDay 2026

## Executive summary

This session introduces GPT Live, a full duplex voice model designed for building complex, interactive voice applications. Key architectural concepts covered include full duplex streaming (allowing simultaneous listening and speaking), robust delegation for complex back-end tasks, and a turnless architecture that maintains functionality even with loud background noise. Developers learn how to integrate GPT Live into existing systems using both client and responses delegation modes.

## Key takeaways

- Full Duplex Voice Model: GPT Live is a full duplex model, meaning it constantly streams audio in and out, allowing it to listen and speak simultaneously. Unlike models reliant on Voice Activity Detection (VAD), it handles turn-taking internally, enabling natural, overlapping conversation flow.
- Delegation for Complex Tasks: For complex requests (e.g., running a game or fetching data), GPT Live delegates work to a separate back-end agent. This keeps the voice conversation flowing while the backend processes the request, which is crucial for maintaining user engagement.
- Turnless Architecture: The turnless architecture ensures the application remains functional and responsive even when exposed to loud or noisy background audio (e.g., background music, chatter), greatly expanding use cases beyond quiet environments.

## Technical details

- GPT Live Architecture: GPT Live operates with a split architecture (UI, Voice Interface, Backend). Delegation allows the voice model to keep streaming audio while the backend performs work using tools like Astra 6 with code execution.
- Delegation Methods: Two primary delegation methods are available: 1) **Client Delegation:** Ideal for existing agent harnesses, where the application receives delegation events and runs the work. 2) **Responses Delegation:** OpenAI manages the delegation process, which is highly familiar to users of tool calling and structured outputs.
- Information Streaming: The backend can provide three types of information to the voice model to maintain context and steer the conversation: 1) **Silent Context Updates:** Keeps the model informed of background state. 2) **Speakable Content:** Passes data through the commentary stream for the model to announce aloud. 3) **Instructions Stream:** Used to change model behavior (e.g., guardrails or system instructions).

## Practical implications

- Build engineers can integrate voice interfaces into existing agent harnesses and workflows, minimizing migration friction.
- The ability to manage background state updates (silent context) allows for complex, multi-step interactions without interrupting the user's conversation flow.
- The turnless architecture makes voice applications viable for public or noisy environments, significantly broadening the potential user base.

## Topics

Voice AI, Full Duplex Streaming, Agent Delegation, Conversational AI, System Architecture, OpenAI Developer plugin, GPT Live Starter app

Source: https://www.youtube.com/watch?v=jDxKzlGLCas
