GPT-Live: full duplex is the feature, decoupled reasoning is the architecture

On July 8, 2026, OpenAI replaced ChatGPT’s Advanced Voice Mode with the GPT-Live family: GPT-Live-1 for paid users and GPT-Live-1 mini for free users. The interesting part is not a better speech synthesizer. It is two system-design decisions — continuous full-duplex interaction, and a strict separation between the interaction layer and the reasoning layer — that together change what a voice interface is.

This note works from the public record: OpenAI’s GPT-Live system card and Realtime API documentation (read directly), the launch post as reflected in consistent independent reporting, and third-party measurements. Where a detail rests only on secondary sources, it is labeled.

Three generations of ChatGPT voice

Generation	Architecture	Turn handling
Standard Voice	Speech-to-text → text model → text-to-speech	Strict sequential turns
Advanced Voice Mode	One audio-native model processes and speaks	Waits for an end-of-turn signal, generally silence
GPT-Live	Consumes input while simultaneously producing output	Continuously decides whether to listen, speak, pause, or delegate

The first system cascaded three models per utterance. Advanced Voice Mode processed audio natively but still modeled a conversation as alternating messages. GPT-Live models the conversation as an ongoing temporal stream: per OpenAI’s launch materials, it makes interaction decisions many times per second — speak, keep listening, pause, interrupt, or invoke a tool. The system card states the models “can follow pauses, interruptions, and changes in pace, and decide in the moment whether to respond or keep listening.”

Full duplex is more than barge-in

Conventional voice agents already support barge-in, and OpenAI’s Realtime API documents that mechanism precisely: when voice-activity detection fires during playback, the server cancels the in-progress response and the client truncates the unplayed audio. (The API’s default turn detection is now semantic_vad — a model-based end-of-turn estimator, not a silence timer — but it is still turn-oriented: user speech ends the assistant’s turn.)

Full duplex removes the assumption that user speech terminates the assistant’s turn. Overlapping input can instead mean: finish the phrase and yield, stop immediately, emit a brief acknowledgment and keep listening, wait out a hesitation, or revise the in-progress response. Backchannels (“mhmm”, “got it”) become a distinct interaction state — producing audio without claiming the floor — rather than a turn. Live translation is the clearest demonstration: there is no clean end-of-turn boundary to wait for, so the system must incrementally understand an evolving stream while producing aligned output.

The two-plane architecture

GPT-Live is not presented as OpenAI’s most capable model. It is a low-latency interaction plane that delegates to a high-capability work plane: the system card confirms both variants “can delegate more complex work to our other models,” and OpenAI’s help materials describe the live model continuing the conversation while a frontier model works in the background and returns its result. Independent reporting consistently identifies the backend as GPT-5.5 (Instant for the fast paths, Thinking for deeper effort); the exact variant naming — GPT-Live-1 Instant / Medium / High mapped to GPT-5.5 effort levels — circulates only in secondary write-ups and should be treated as unverified.

The separation matters more than any single capability. One model no longer has to optimize simultaneously for time-to-first-audio, prosody, turn-taking, noise robustness, deep reasoning, and long-running tool work. And the planes upgrade independently: OpenAI can swap the reasoning backend without retraining the conversational front-end.

Safety also moved into the stream. The system card describes safeguards that check inputs and generated outputs “as the conversation unfolds” and can steer or interrupt a response, play a spoken safety message, provide support resources in text, or end the session — moderation inside the generation loop, not after a finished message.

What independent measurement adds

OpenAI published benchmark deltas — reported as GPQA 84.2% vs 45.3% and BrowseComp 75.2% vs 0.7% against Advanced Voice Mode, with 75.7% (GPT-Live-1) and 69.2% (mini) human preference — but analysts note those reasoning scores largely measure the delegated GPT-5.5 backend, not the voice layer, and all comparisons are against OpenAI’s own predecessor.

The most useful independent data comes from Agora, which measured GPT-Live end-to-end on a real device in launch week because OpenAI published no latency figures. Their findings reframe the improvement: median response latency of roughly 1.1 s (vs ~1.3 s for Advanced Voice Mode) — not dramatically faster, but far more consistent (P90 barely above median, where Advanced Voice Mode’s P90 exceeded 2.3 s). GPT-Live rejected all 30 of 30 background-speech false-interruption probes where Advanced Voice Mode stopped for 20. The cost: it takes about half a second longer to go silent when genuinely interrupted. Field reports match that tradeoff — early users found the backchannels overeager (including a reported bug where it laughed and interjected mid-sentence), apparently tuned down after launch. The GPT-4o-era “232 ms minimum / 320 ms average” figure belongs to a 2024 announcement and should not be carried forward to GPT-Live.

Prior art, and what is actually new

Neither half of the design is novel as research. Kyutai’s open-source Moshi (2024) demonstrated real-time full-duplex spoken dialogue with a neural audio codec; Alibaba’s Qwen-Omni line split a “thinker” from a “talker”; Nvidia’s PersonaPlex (2026) ships open weights built on Moshi; academic work from dGSLM onward covers duplex modeling, and Full-Duplex-Bench now evaluates overlap handling directly. What is new is the integration at consumer scale: continuous duplex interaction, a model-driven conversational control policy, stream-embedded safety, and asynchronous delegation to a frontier reasoner, shipped as the default voice interface for a very large user base.

What this means for self-hosted voice

The delegation pattern is the replicable idea. A local equivalent does not require replicating GPT-Live’s model: an open full-duplex front-end (Moshi or a derivative) can proxy to any backend reasoner — local or remote — over an internal channel, which is precisely the shape the open-source ecosystem is already exploring. The hard parts, judging by the measurements and field complaints above, are not synthesis quality but interaction tuning: backchannel calibration, yield latency on interruption, echo handling while the system’s own output is audible, and tool calls that never block the audio loop. Those are operational contracts — the same category of problem as readiness gating and streaming health semantics — and they are where a self-hosted duplex stack will be won or lost.

Limitations

This note documents third-party systems from public materials retrieved on 2026-07-21; nothing here is based on access to OpenAI infrastructure. OpenAI’s launch post and Help Center refused automated retrieval, so several attributions rest on multiple consistent secondary reports rather than primary text, as labeled. The Instant/Medium/High variant naming and an “internal voice-telecom benchmark” mentioned in some write-ups remain unverified. Agora’s measurements are a single-device, single-location, launch-week snapshot. OpenAI has published no parameter counts, tokenizer or codec details, streaming-state design, or a technical paper for GPT-Live; architectural statements above are system-level disclosures, not implementation facts.