OpenAI has released GPT-Live-1 through its API, putting the full-duplex voice model behind ChatGPT’s voice mode into developers’ hands at $0.05 per minute, billed per second.
The announcement landed Sept. 10 on the OpenAI developer forum and the company’s site. GPT-Live-1 listens and speaks at the same time rather than taking turns, handles interruptions natively, and manages backchannels, the “mhmm” and “yeah” sounds that make a conversation feel alive. Deeper reasoning and tool use are delegated to a backend model of the developer’s choice, keeping the voice layer fast while the heavy thinking happens elsewhere.
Until now, developers who wanted OpenAI voice had to chain separate systems: speech recognition, a language model, then synthesis. Each boundary added latency and each one was a place things broke. Interruption was the clearest failure case. Cutting off a turn-based voice assistant mid-sentence produces the stilted, walkie-talkie feel that has defined most voice products to date. A full-duplex architecture treats interruption as a normal part of conversation instead of an error state.
The model was first rolled out to ChatGPT users on July 8 alongside a smaller variant, GPT-Live-1 mini. At launch, OpenAI said complex questions would be delegated to GPT-5.5 behind the scenes. The API version generalizes that: developers can connect any backend model, agent framework or tool set, and the voice layer handles the live conversation while delegated work runs in the background. Interrupting the speech does not cancel the backend task, which matters for anything longer than a quick lookup.
Audio generated through GPT-Live carries SynthID watermarking, introduced at the end of July, and OpenAI exposes an API for verifying provenance signals in supported files. That gives call-center operators and app builders a way to distinguish model-generated speech from human recordings, a capability that is becoming relevant as voice cloning complaints pile up in state legislatures.
The benchmarks and what they measure
OpenAI’s published numbers focus on task completion through voice. Paired with GPT-6 Astra at medium reasoning effort, GPT-Live-1 completed 83.6 percent of Tau3 tasks on the first attempt, against 45.7 percent for the previous GPT-Realtime-2.1. Tau3 covers airline, retail and telecom support scenarios. On TauBanking, which tests document retrieval and account tool use, the pairing scored 38.1 percent.
| Metric | GPT-Live-1 | GPT-Realtime-2.1 |
|---|---|---|
| Tau3 first-attempt completion | 83.6% | 45.7% |
| Turn-taking latency | 0.798s | 1.41s |
| Full Duplex Bench v1.5 interactivity | 80.1% | not disclosed |
| Tool-calling success (v3) | 87% | not disclosed |
Turn-taking latency dropped to 0.798 seconds from 1.41 seconds, which is the difference between a conversation and a series of exchanges. Third-party reviewers note these figures have not been independently reproduced, and the exact conditions under which the model hands off to a backend model remain undocumented. The independent review from kie.ai rated the interaction-design evidence as the strongest part of the release and flagged the delegation handoff as the biggest open question.
Telephony support is included, which points at the commercial target: customer service. Voice support is the largest installed base of conversational AI spending, and a per-second-billed voice model with native interruption handling is aimed directly at call centers that currently run brittle IVR trees or expensive human agents. Integration with the Codex SDK lets a voice session delegate repository investigation or other long-running work while the caller keeps talking.
The competitive picture
OpenAI is not alone in full-duplex voice. Google’s Gemini Live voice modes reached Gmail, Docs and Keep for paying subscribers this week, extending voice interaction into document work rather than standalone chat. ElevenLabs and a range of smaller vendors serve the same market with stacked pipelines. The differentiator OpenAI is claiming is the single model that handles listening, speaking and turn management natively, with delegation as a first-class API concept rather than a developer-side hack.
Pricing matters in this market. At $0.05 per minute, a continuous eight-hour voice agent costs $24 per day per line before backend model costs, which are billed separately. That is cheap enough for high-volume support lines but not free, and developers building consumer apps will watch how the bill scales with talkative users. Backend usage is the wild card: a voice session that triggers frequent GPT-6 Astra calls can cost several times the voice fee itself.
The July rollout included live translation and background-noise handling, both carried into the API version. Neither is unique to OpenAI, but combined with sub-second turn-taking they form a package that is hard to replicate with chained components, because every added pipeline stage adds its own delay.
The release also arrives as OpenAI’s consumer and enterprise businesses pull in different directions. The company has been cutting prices on some text models while shipping premium-priced capabilities like this one, and voice is one of the few categories where customers have shown willingness to pay per minute rather than per token. If GPT-Live-1 becomes the default substrate for phone-based agents, the per-minute meter becomes a significant revenue line.
For ChatGPT users, nothing changes immediately. The consumer voice experience already runs on these models. What changes is who can build on them. Voice agents for booking, support, education and accessibility have been demo projects for years because the pipeline approach made them feel robotic. If the full-duplex numbers hold up outside the benchmark suite, the constraint moves from technology to design: what should a voice agent actually be allowed to do, and who is liable when it does it.
Developers can start with the launch post and quickstart on the OpenAI site. The API supports custom voices, stronger instruction following than the ChatGPT consumer experience, and both managed delegation to OpenAI models and client-side delegation to external services.