OpenAI Has Been Holding Back GPT-Bidi-1 for Half a Year — Aiming to Teach Voice Mode to Handle Interruptions

On June 17, testingcatalog released a leak: OpenAI is preparing a major update for ChatGPT’s voice mode, with a new model codenamed GPT-Bidi-1 now entering the rollout preparation phase. This is the most substantial voice architecture upgrade since the launch of Advanced Voice Mode in 2024 — not just a change of voice or addition of emotion, but a fundamental shift in the underlying conversation mechanism from “taking turns” to “bidirectional parallel processing.”

This has been foreshadowed for a while. Earlier this year, The Information reported that OpenAI was working on a BiDi architecture, originally planned for release in Q1. But prototypes started speaking nonsense or changing voice after a few minutes of continuous dialogue, dragging the schedule all the way to mid-year. Now we’re finally seeing signs of release — not too late, considering text models have already moved up to GPT-5.5 while voice is still stuck on last year’s stack, making the gap increasingly awkward.

Where Exactly Does “Taking Turns” Get Stuck?

To understand the significance of BiDi, you first need to know how Advanced Voice Mode currently works. Essentially, it’s still a half-duplex system: you finish speaking, the model’s VAD (Voice Activity Detection) decides you’ve stopped, then starts processing the audio, generates a reply, and plays it back. If you suddenly interject with “wait” or “that’s not what I meant,” the model either forcibly interrupts or finishes what it planned to say before switching back to listen.

This mode has several inherent flaws:

• Interruptions feel unnatural. In human conversation, backchannel feedback like “mm-hmm,” “right,” or “and then?” is common, but the current model hears any sound and thinks you are taking over, instantly shutting up.
• Rigid pacing. AI can’t adjust its response direction mid-sentence; it has to wait for you to finish, think, and then output all at once.
• Heavy perception of latency. Even if end-to-end latency is reduced to 300ms, it’s still a three-step “wait-think-speak” process — far from the parallel feel of humans who listen and organize their response at the same time.

GPT-Bidi-1 aims to solve exactly this. The core of bidirectional is keeping both input and output streams open at all times, so the model can continue listening to the user even while speaking, and can modify subsequent content in real time based on new input. In other words, it’s no longer “listen to a segment, then speak a segment,” but “constant listening, constant speaking, and changing speech according to what’s heard.”

This is True “Full Duplex”

Anyone who has built real-time audio systems knows full-duplex conversation sounds simple, but is an engineering deep pit. The model has to simultaneously maintain two states: what it is saying, and what the user is saying — and decide within milliseconds whether the user’s utterance is feedback, supplement, interruption, or noise. Misjudging can cause it to treat “mm-hmm” as a new command or treat a real interruption as agreement.

From leaked details, GPT-Bidi-1’s abilities include:

1. Continuous audio stream processing. No more waiting for VAD to trigger — it’s constantly parsing input, deciding on its own when to yield or when to continue speaking.
2. Mid-course directional adjustment. For example: you tell an AI customer service rep “I want to return an item” and halfway through you change your mind to “actually, exchange it instead.” The old model would stall or still execute the return process, whereas BiDi could theoretically switch branches seamlessly.
3. Smoother tool invocation. OpenAI has previously revealed that BiDi performs significantly better when chaining external tools (function calling), which is essential for voice agents interfacing with business systems.

OpenAI internally sees this as the key leap to push AI assistants from “toy” to truly deployable customer service and hardware interaction. They’re also planning voice-first hardware — smart speakers, wearable devices — where half-duplex experience simply can’t hold up.

High / Medium / Instant Three Levels of Intelligence

The product form after GPT-Bidi-1’s launch is also noteworthy. Users can freely switch between bidirectional mode and Advanced Voice Mode — a pragmatic choice, since BiDi is more natural but the “always listening, always speaking” overhead is obviously higher than turn-based mode, and not every scenario needs it.

Interesting as well is that it supports three intelligence levels:

• High: For complex tasks and reasoning-heavy conversations, slower responses but higher depth.
• Medium: Balanced mode for daily Q&A and information retrieval.
• Instant: Prioritizes quick reactions, suited for backchannel-heavy small talk or control commands.

This tiered approach aligns with OpenAI’s current reasoning_effort for text — exposing reasoning depth for users to choose. For developers, this means you can tune according to SLA and cost budget — e.g., voice customer service can default to Medium, switch to High for critical points, and run simple tasks like restaurant booking directly in Instant.

Comparing with Gemini, Sesame, and Others

Looking across the industry, full-duplex voice is not just OpenAI’s pursuit. Google’s Gemini Live, Sesame’s CSM, and several domestic companies working on on-device voice solutions each have slightly different approaches but share the same goal: making AI sound like a human.

Gemini Live takes a native audio route with end-to-end audio modeling, which has good rhythm and emotion handling but conservative interruption handling. Sesame’s voice presence on small models is impressive, but scaling and tool usage capability lag behind. If OpenAI succeeds in stabilizing BiDi, its biggest advantage is combining full-duplex conversation + GPT-level reasoning + tool invocation in one package — others have attempted the first two, but combining all three effectively for commercial use has not yet been seen.

Of course, “being ready to launch” and “being usable” are not the same. Prototype crashes and voice changes after a few minutes suggest that long-context audio modeling stability is still being fought over. Likely, the initial release will be a limited-duration beta to gather data.

What This Means for Developers

If you’re working on voice-related products, here are points worth considering now:

• Leave switch space in your existing Realtime API code. When BiDi launches, it will likely bring new interfaces or parameters, potentially rendering half-duplex turn detection logic obsolete.
• Adjust prompt design. In half-duplex times, you could assume “the user finishes a full sentence”; with BiDi, the user might change their mind in the third second of your answer, so your system prompt must explicitly handle mid-conversation changes.
• Customer service, education, and companion scenarios will be redefined. Voice bots that were barely acceptable before may become outdated after BiDi arrives — evaluate upgrade paths early.

As a side note, OpenAI Hub (openai-hub.com) will integrate GPT-Bidi-1’s API as soon as it’s officially available, supporting direct access in China, allowing a single key to call GPT / Claude / Gemini / DeepSeek in OpenAI-compatible format. If you want to run a bidirectional voice demo when the new model launches, invocation is expected to be similar to the current Realtime API:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.openai-hub.com/v1",
    api_key="sk-your-hub-key"
)

# Expected interface shape (subject to official release)
session = client.beta.realtime.sessions.create(
    model="gpt-bidi-1",
    modalities=["audio", "text"],
    voice="alloy",
    turn_detection={
        "type": "bidi",          # Enable bidirectional mode
        "intelligence": "medium" # high / medium / instant
    },
    instructions="You are a retail customer service agent who must seamlessly switch tasks when the user changes their mind mid-conversation."
)

print(session.client_secret.value)

The field names above are guessed based on the current Realtime API style; final details will depend on OpenAI’s documentation. But the concept is close: change turn_detection from server_vad to bidi and add an intelligence level.

In Closing

Voice is one of the most underestimated yet hardest-to-get-right parts of AI interaction. No matter how smart text dialogue gets, its experience ceiling is just “typing speed”; but if voice can reach human-level quality, AI assistant penetration will jump another tier — and OpenAI sees this clearly, choosing to delay half a year to refine BiDi.

Whether GPT-Bidi-1 can truly solve the “mechanical feel” problem will only be known after a few rounds of testing. But one thing is certain: the era of turn-based voice AI is ending, and the voice agent track will heat up in the second half of the year.

References

• ITHome - ChatGPT’s Biggest Voice Upgrade: OpenAI Preparing to Launch GPT-Bidi-1 AI Model: Original leak report, detailing GPT-Bidi-1’s name and three-tier intelligence levels.