OpenAI Builds Bidi 1 Into ChatGPT: A Voice Model That Talks and Listens at the Same Time

OpenAI is accelerating ChatGPT's transformation into a super app: the company is developing the bidirectional voice model Bidi 1, building in the Codex coding assistant and expanding agent capabilities, all at once.

What is GPT Bidi 1

Researchers found traces of a new audio model, codenamed GPT Bidi 1, in the code of the ChatGPT mobile app. "Bidi" stands for bidirectional.

This is a fundamentally different approach from the current voice mode, where the model switches in sequence between listening and answering: first it listens, then it speaks, like a walkie-talkie with a push-to-talk button. Duplex mode changes the dynamics of the conversation itself. Users will be able to interrupt the assistant mid-sentence, and the assistant will be able to react to intonation, pauses and speech tempo in real time, without waiting for a pause. The delay between turns drops noticeably, and the conversation feels less robotic.

This is how the most advanced voice interfaces work: Apple Intelligence in iOS 18 and Google Duplex have long supported full duplex exchange.

What changes with Bidi 1:

• voice and hearing work simultaneously, with no queue for the "airwaves"
• the user can interrupt a response at any moment
• response time between turns gets shorter
• the assistant hears intonation and reacts to pauses
• the conversation sounds more natural and closer to talking with a person

No specific release date for Bidi 1 has been announced yet, but the model's presence in ChatGPT's production code suggests the release is near.

Codex inside ChatGPT

Alongside its voice work, OpenAI is building its programmer's assistant, Codex, into ChatGPT. Until now, Codex has existed as a separate product with its own interface. After the integration, users will be able to write, debug and explain code directly in the main chat window, without switching tabs.

For developers who have already made ChatGPT part of their workflow, this is a significant simplification: one unified workspace replaces two separate tools. Codex will be available in the same conversation as web search, image generation, file analysis and, soon, the Bidi 1 voice mode.

Agents: the third front

The third component of the large update is tooling for AI agents. ChatGPT is steadily evolving from a reactive chatbot into a proactive assistant that can carry out multistep tasks autonomously: booking tickets, comparing products on marketplaces, filling out online forms, and finding and synthesizing information from different sources.

"We want to make

ChatGPT the app that does for you everything you normally do on the internet," say sources familiar with OpenAI's strategy.

All three components (the live voice of Bidi 1, the built-in Codex and autonomous agents) reinforce one another. An agent with duplex voice and the ability to write code is a fundamentally different class of tool, not just an updated chatbot.

What this means

OpenAI is deliberately moving toward the super app format: a single product covering most everyday scenarios, including voice, code, task automation and information search.

For ordinary users, the most immediate consequence is that talking to ChatGPT will become noticeably livelier and more natural. For businesses, it means a new level of automation for operational tasks that previously required a person or a set of separate SaaS tools.

Competitors, from Google to Anthropic, will have to speed up again.