OpenAI shipped GPT-Live-1 on 10 September 2026, a single model that listens and speaks at the same time, rather than the listen-then-think-then-speak pipeline most voice agents run today. The voice layer is priced at $0.05 a minute, with reasoning and tool calls billed separately on top of that.
If you're building, or have been putting off building, a voice-based product — phone support, a voice assistant feature, an outbound calling agent — this is a genuine architecture change, not just a new model to swap into the same pipeline.
What changes about the architecture
The standard way to build a voice agent today chains three separate stages: a speech-to-text model transcribes what the caller said, a language model decides what to say back, and a text-to-speech model turns that decision into audio. Each handoff adds latency, and the pipeline can't easily handle a caller talking over the agent mid-response, because the stages are sequential.
GPT-Live-1 is built to listen and speak simultaneously inside one model, which is what OpenAI means by full-duplex. According to OpenAI's own announcement, it claims a 30% improvement on full-duplex benchmarks compared with GPT-Realtime-2.1, OpenAI's previous realtime voice model. That benchmark figure is OpenAI's own reported number, not yet independently verified by a third party, so treat it as a vendor claim worth testing on your own use case rather than a settled fact.
What the pricing actually means for your bill
The $0.05-per-minute figure covers the voice layer specifically — the listening-and-speaking part of the interaction. Reasoning and tool calls, the part where the model actually decides what to do or look something up, are billed separately according to OpenAI's standard token pricing. That split matters for budgeting: a simple, short voice interaction that mostly involves speaking and listening will cost close to the per-minute rate, while a longer call involving several tool calls or complex reasoning will cost meaningfully more than the headline $0.05 figure suggests. Anyone estimating costs for a voice product should model both components separately rather than quoting the per-minute rate as the full cost.
Why this matters for a small team specifically
Maintaining a three-stage voice pipeline is real engineering overhead: three separate model integrations, three sets of latency to manage, and a genuinely hard problem in handling interruptions gracefully across stage boundaries. A single model that handles both listening and speaking removes an entire integration surface, which is the kind of simplification that matters disproportionately to a small team without dedicated voice-infrastructure engineers.
That doesn't mean every voice product should switch immediately. If you already have a working three-stage pipeline tuned carefully for your specific use case, a lower cost or better benchmark on a new model isn't automatically worth a rebuild. But if you've been putting off building a voice feature specifically because a three-stage pipeline felt like too much infrastructure for the team you have, this removes a real piece of that objection.
What to actually do if you're considering a voice product
Run a real comparison rather than trusting the 30% benchmark figure directly. Take a handful of representative conversations your product would actually need to handle — including at least one where the caller talks over the agent mid-sentence, since that's the specific case full-duplex models are meant to handle better — and test them against both your current pipeline (if you have one) and GPT-Live-1 directly.
Model the full cost, not just the per-minute voice rate. Estimate a typical call's mix of pure conversation time versus tool-call-heavy exchanges, and price out both components separately using OpenAI's published rates, so you're not surprised by a bill that looks nothing like $0.05 a minute once reasoning costs are added in.
If you're currently maintaining a three-stage pipeline that works, don't rebuild it purely because a single-model alternative exists. Rebuild it if the interruption-handling or latency improvements solve a problem your current setup actually has, which is a different bar than "the new thing is newer."
The bottom line
GPT-Live-1, shipped 10 September 2026, replaces the standard three-stage voice pipeline with a single model that listens and speaks simultaneously, priced at $0.05 a minute for the voice layer with reasoning billed separately. If a voice feature has been on your roadmap but felt like too much infrastructure, test this against a real conversation sample, including interruptions, before deciding whether it changes that calculation.