GPT-Bidi-1: Everything We Know About OpenAI's Bidirectional Voice Model for ChatGPT

OpenAI is preparing its biggest voice-mode upgrade for ChatGPT yet: a bidirectional audio model internally called GPT-Bidi-1. Unlike every previous ChatGPT voice mode, which waits for a user to finish speaking before responding, GPT-Bidi-1 can listen and talk at the same time — absorbing interruptions, correcting mistakes in real time, and adjusting its replies without dropping the audio track.

This article covers how GPT-Bidi-1 was discovered, what makes it different from turn-based voice, the three intelligence tiers it introduces, the visual change to the ChatGPT interface, and the competitive context of its development. The model has not been officially announced by OpenAI, so all details come from code sightings, UI references, user reports, and media analysis.

How GPT-Bidi-1 was discovered

The discovery chain started with developer M1Astra, who first spotted references to gpt-bidi-1 in ChatGPT's app code and shared the finding on X. The tracking site TestingCatalog then confirmed the model string, alongside what appeared to be announcement text describing "the next generation of Voice" and a "major leap in intelligence."

Code and UI elements were subsequently found on both web and mobile ChatGPT clients. Limited tests began flowing to a small subset of users in late June 2026. By June 22–24, 2026, multiple user reports and demonstration videos emerged showing the model working bidirectionally in practice.

Caveat: OpenAI has not issued an official announcement. The model's final name, exact tier behavior, and rollout date remain unconfirmed by the company.

How bidirectional voice differs from turn-based models

Current ChatGPT voice modes — Standard Voice and Advanced Voice Mode — operate in a turn-based paradigm. The model must wait for the user to finish speaking before it can respond. GPT-Bidi-1's bidirectional (BiDi) architecture allows the model to process two audio streams simultaneously: yours and its own.

Key behavioral differences reported in demonstrations:

• Interruption handling: The model can absorb interruptions mid-sentence and dynamically adjust its response without freezing or dropping the audio track.
• Real-time correction: In tests, GPT-Bidi-1 counted alongside a user from 1 to 10, then immediately switched to counting backwards when the user interrupted mid-stream.
• Natural overlap: The assistant can speak while the user is still talking, making exchanges closer to human conversation.

OpenAI's internal goal was to close the gap between ChatGPT's voice stack — which lagged behind its text models (already at GPT-5.5-class reasoning) — and deliver parity in real-time conversational intelligence.

Three selectable intelligence tiers

GPT-Bidi-1 is the first OpenAI voice model to introduce three selectable intelligence and speed tiers for voice:

Tier	Description
High	Maximum reasoning depth, slower response — for complex analysis tasks
Medium	Balanced trade-off between intelligence and speed
Instant	Fastest possible response, reduced reasoning — for casual or time-sensitive interactions

The tier system lets users tailor interaction depth versus latency per task, similar to how ChatGPT's text models offer different reasoning levels. For example, a quick weather query would use Instant, while a deep brainstorming session would switch to High.

Visual changes to the voice interface

When GPT-Bidi-1 is selected, the voice bubble/waveform indicator changes to yellow instead of the current default color. The model appears in the settings model-selector as a new option labeled "Bidi (Latest)" alongside existing Standard Voice and Advanced Voice Mode, rather than replacing them.

Development timeline and competitive context

• Early 2026: OpenAI began developing the BiDi architecture internally.
• March 5, 2026: The Information reported that OpenAI was developing a bidirectional audio model to boost its voice assistants.
• May 7, 2026: OpenAI announced GPT-Realtime-2 in the API — its first voice model with GPT-5-class reasoning, which laid groundwork for bidirectional capabilities.
• June 16–17, 2026: M1Astra and TestingCatalog first publicly spotted code references for gpt-bidi-1.
• June 22–24, 2026: User reports and demonstration videos emerged showing working bidirectional behavior. Multiple outlets reported that a wider rollout might come within the same week.

Competitive context: The bidirectional voice push directly responds to advances from Google (Gemini Live with interruptions), Anthropic, and real-time voice agents from startups. OpenAI is racing to bring voice interaction parity to its text intelligence, which already powers GPT-5.5-level reasoning.