OpenAI Releases GPT-Live Full Duplex Voice Model

OpenAI released GPT-Live, a full duplex voice model that listens and speaks simultaneously. The system handles interruptions, real time feedback, and complex tasks by routing them to a background GPT-5.5 instance.

The launch makes two versions available immediately to ChatGPT users worldwide. Human evaluations over five to ten minute conversations showed gains over Advanced Voice Mode in natural flow, turn taking, and interruption handling. Benchmarks such as GPQA, BrowseComp, and τ3-Voice Telecom also recorded stronger results.

Model Architecture and Real Time Decisions

GPT-Live runs on a full duplex design that checks multiple times per second whether the user is speaking, listening, interrupting, or triggering a tool call. Short latency decisions stay inside the model. Longer operations such as search or reasoning shift to the background GPT-5.5 instance so the main voice channel stays uninterrupted.

This split keeps the spoken exchange continuous while still allowing access to heavier computation when needed. The company states the architecture supports natural pauses and quick corrections without forcing the user to wait for a full response cycle to finish.

Performance Claims in Benchmarks and Human Tests

Evaluations compared GPT-Live-1 and GPT-Live-1 mini against the prior Advanced Voice Mode. Reviewers rated the new models higher on naturalness, conversation rhythm, and handling of mid sentence interruptions. The gains appeared consistently across five to ten minute sessions.

On GPQA, BrowseComp, and τ3-Voice Telecom the models posted higher scores than earlier voice systems. OpenAI attributed part of the improvement to cleaner separation between real time voice control and deeper reasoning tasks.

Immediate Availability and API Plans

Both GPT-Live-1 and the smaller GPT-Live-1 mini version opened to all ChatGPT users on the day of the announcement. No wait list or regional restrictions were mentioned for the initial rollout.

The company also said it plans to release an API version later. No timeline was given, but the statement positioned the API step as the next extension beyond the current ChatGPT integration.

Comparison With Earlier Voice Systems

Advanced Voice Mode already allowed spoken conversation inside ChatGPT. GPT-Live changes the underlying structure by allowing true overlap of input and output. Users no longer need to finish a full turn before the model can respond or adapt.

The background delegation mechanism is another difference. Earlier voice features kept most processing inside a single model instance. Routing heavy work to GPT-5.5 frees the voice layer to stay responsive during complex queries.

Limits and Open Questions

The current release covers only ChatGPT users. Enterprise or developer access still depends on the future API release. How the delegation layer will be priced or rate limited remains unclear.

Human evaluation scores cover conversations of five to ten minutes. Performance over longer sessions or in noisy environments has not been detailed. Independent tests will be needed to confirm whether the reported gains hold outside the controlled conditions described by OpenAI.

What to Watch Next

Developers will look for the API announcement and any documentation on latency, rate limits, and access tiers. Usage growth inside ChatGPT will show whether the new interruption style changes daily behavior for a meaningful share of users.

Competitors that offer voice agents will likely release updated models or public benchmarks within the next quarter. Any side by side comparisons on shared test sets will clarify whether the full duplex approach delivers durable advantages or remains tied to one vendor stack.