OpenAI’s newest voice model separates listening and speaking from thinking, at least in price. GPT-Live-1, now live in the API, costs $0.05 per minute; whichever reasoning model and agent harness handle the actual answers, be it GPT-6 Astra, Codex, or ChatGPT Work, gets billed on top.
Instead of routing a call through transcription, a separate reasoning pass, then synthesized speech, the model takes in and sends out audio at once. OpenAI says this removes the latency and lost context that plague that older, chained setup, letting the model acknowledge and interrupt naturally rather than narrate each step.
Early adopters include Yelp Host, Intercom’s Fin, and Cognition’s Devin. Speak, the language-learning app, told OpenAI its early tests showed nearly 80% fewer interruptions than its prior turn-based system, a company-reported figure rather than an independent benchmark. One customer said switching off a cascaded pipeline cut 23,000 lines from its voice codebase, an 80% reduction, on a system built for real-time patient conversations, per OpenAI.
Twelve voices ship at launch, with custom voices reserved for sales conversations. For any team still running a chained speech pipeline, that per-minute price is a concrete number to weigh against the engineering cost of keeping one running.
TestingCatalog reported the GPT-Live-1 launch on September 10, 2026.
