OpenAI has begun rolling out gpt-bidi-1, a next-generation voice model for the ChatGPT app that enables full-duplex, bidirectional communication — meaning the AI can listen and speak at the same time. Early users are already sharing demos where ChatGPT interrupts them mid-sentence, counts along as they speak, sings on request, and corrects pronunciation errors in real time.
Key Highlights
- gpt-bidi-1 is a bidirectional audio model that listens and speaks simultaneously
- The model can interrupt users mid-sentence and accept interruptions without freezing
- Three intelligence tiers introduced: Instant, Medium, and High
- Rolling out to a subset of ChatGPT app users as of June 21–22, 2026
- Users will toggle between the new "Bidi (Latest)" mode and the existing Advanced Voice Mode
What Makes It Different
Current ChatGPT voice (Advanced Voice Mode) operates like a walkie-talkie: the model freezes the moment a user speaks while it is responding. GPT-Bidi-1 eliminates that bottleneck with a bidirectional architecture that processes incoming and outgoing audio at all times.
Early demos shared on X show the model:
- Counting out loud while a user counts in parallel
- Correcting a user's word mid-sentence without missing a beat
- Singing full songs when prompted — a capability users confirmed was absent before the update
- Absorbing "mm-hm" and other interjections naturally, without breaking conversation flow
"It's quite impressive. The AI can listen and speak at the same time, making the conversation much more natural," wrote developer Mark Kretschmann after testing the rollout.
Three Intelligence Tiers
GPT-Bidi-1 introduces a tiered system that mirrors the text-side model lineup:
| Tier | Description |
|---|---|
| Instant | Lowest latency, suitable for quick back-and-forth |
| Medium | Balanced speed and reasoning |
| High | Deepest reasoning, slower response |
This alignment brings voice conversations to the same capability level as GPT-5.5-era text interactions for the first time.
Why It Matters
OpenAI's text models have advanced rapidly across 2025–2026, while the voice stack lagged behind — creating a noticeably weaker experience when talking versus typing. GPT-Bidi-1 is designed to close that gap directly.
The update is already being compared to the "Her" moment — a reference to Spike Jonze's 2013 film about AI as a natural conversational partner. The combination of gpt-bidi-1 with OpenAI's existing Codex agents and computer-use capabilities points toward a future where users direct their entire computer through voice alone.
The timing is also strategic. ChatGPT's global market share recently dropped below 50% for the first time as competitors including Google Gemini and Claude improved their own voice and multimodal capabilities. A full-duplex voice mode is a significant differentiator at a moment when the AI assistant market is increasingly contested.
Rollout Status
The rollout is gradual. As of June 22, 2026, a subset of ChatGPT app users on web and mobile are seeing the new voice mode. OpenAI has not published an official announcement, but app activity, model name sightings (gpt-bidi-1), and user-shared demos confirm the release is live. The final product name may differ from the current internal tag.
Other OpenAI models spotted in parallel testing include ember-alpha, beacon-alpha, and early builds of GPT-5.6 — suggesting a broader model refresh is underway.
Background
OpenAI launched gpt-realtime-2 in early 2026 as a developer API for building voice agents, with streaming audio and tool-call support. GPT-Bidi-1 is the consumer-facing evolution of that same bidirectional architecture, bringing the capability directly to the ChatGPT app without requiring any developer integration.
What's Next
- Full rollout to all ChatGPT users is expected in the coming days
- GPT-5.6 (including
ember-alphaandbeacon-alphavariants) is also in late-stage preparation - OpenAI is expected to position bidirectional voice as a core differentiator ahead of its planned IPO
Source: Testing Catalog