OpenAI Brings GPT-Live-1 to the API at 5 Cents a Minute

OpenAI began selling access to GPT-Live-1, its full-duplex voice model, through the API on Sept. 10 at $0.05 a minute for the voice layer alone.

The model has powered ChatGPT Voice since July 8, when OpenAI rolled out GPT-Live-1 and a smaller GPT-Live-1 mini to ChatGPT users worldwide, OpenAI said in its announcement. The Sept. 10 release opens the same model to developers building separate voice products, rather than introducing a new one.

GPT-Live-1 listens and speaks at the same time instead of chaining separate speech-to-text, reasoning and text-to-speech models, and callers can interrupt without breaking the exchange. OpenAI said the model hands deeper reasoning to a backend model of the developer's choice, including GPT-5.6 Luna, GPT-6 Astra or a third-party system, billed separately from the voice fee.

The latency and accuracy numbers

Metric	GPT-Live-1	GPT-Realtime-2.1
Tau3 Voice Intelligence (Pass@1)	86.2%	45.7%
Tau Banking Voice Knowledge	32.0%	12.4%
Turn-taking latency	0.798s	1.41s

The figures above are OpenAI's own, run on its Full Duplex Bench and two Tau benchmarks; The Terminal could not independently verify them. OpenAI said the model ships with 12 voices at launch and that concurrent sessions scale from 25 on the lowest usage tier to 500 on the highest.

Speak, a language-tutoring app, said the model cut interruptions during learners' pauses to think by close to 80 percent compared with its earlier turn-based system, according to OpenAI's announcement. Andrew Hsu, Speak's co-founder and chief technology officer, is quoted in the same post describing the change.

OpenAI's announcement also quotes Alex Levy, Yelp's chief technology officer, saying callers speak fuller, more natural sentences on the new model, and Jordan Neil, chief operating officer at customer-support company Fin, saying the shift moves automated phone support closer to the flow of an ordinary call. Walden Yan, co-founder and chief product officer at Cognition, said his team pairs GPT-Live-1 with the coding agent Devin for voice-driven handoffs between a developer and the agent, though OpenAI did not say how many of these integrations are in production versus still in testing.

What the API version drops

GPT-Live-1 does not accept image input in the API, a capability its predecessor GPT-Realtime-2.1 had. OpenAI did not say whether or when that capability will return. Its training data carries a knowledge cutoff of July 31, 2025, according to OpenAI's documentation.

OpenAI has not published a worked example of what a finished voice agent costs per call, the number developers need to budget with, since the backend model, telephony and any tools are billed on top of the $0.05-a-minute voice fee, DataNorth AI wrote in its own review of the launch. The gap matters more for the Agents API OpenAI opened the same week, which is meant to run GPT-Live-1 alongside other billed components.

GPT-Live-1's release follows OpenAI's push into enterprise-specific products this month, aimed at giving specific industries a reason to route more of their work through OpenAI's API rather than a competitor's. OpenAI said it plans to add more languages to GPT-Live-1. It did not say when.