OpenAI made its GPT-Live-1 voice model available to developers through the API on September 10, 2026, according to The Decoder. OpenAI's own documentation describes it as a full-duplex voice model for real-time conversations, meaning it can listen and speak simultaneously, and calls it the company's premier model for natural, expressive voice conversations with smooth interruption handling.
According to The Decoder, OpenAI reported that GPT-Live-1 reaches an 80.1 percent score on full-duplex interactivity tests, against 45.4 percent for GPT-Realtime-2.1, and cuts turn-taking latency to 0.8 seconds from 1.4 seconds. Tool-calling accuracy rose to 87 percent from 60 percent, and the pass rate on a banking voice-support benchmark rose to 32 percent from 12.4 percent. All of those figures were reported by OpenAI and have not been independently verified. The model ships with 12 new voices covering additional accents, dialects and languages, and can delegate heavier reasoning to backend models, which are billed separately. Sessions include automatic speech-recognition transcripts alongside response text.
Pricing and technical limits
OpenAI's developer documentation prices voice sessions at $0.05 per minute, billed per second, and states that session duration is not rounded up to the next whole minute; backend model and tool usage are charged on top. Access is limited to the Live endpoint (v1/live/sessions). The documentation lists Chat Completions, Responses, Realtime, Realtime translation and transcription, Assistants, Batch, fine-tuning, embeddings, image generation and editing, video, speech generation, transcription, translation, moderation and legacy Completions as unsupported. Concurrent session limits scale by paid usage tier: 25 sessions on Tier 1, 50 on Tier 2, 200 on Tier 3, 300 on Tier 4 and 500 on Tier 5. Free-tier accounts cannot use the model, which also does not support image or video input, structured outputs or predicted outputs. The documentation lists streaming and function calling as supported features, and gives the model a knowledge cutoff of July 31, 2025.
Early business use
The Decoder reported that Yelp is already using GPT-Live-1 for phone-based reservations, with the company's chief technology officer, Alex Levy, reporting improved call handling. Yelp's deployment illustrates the pattern OpenAI is pitching to other businesses: GPT-Live-1 manages the live conversation while heavier reasoning or lookups are routed to other models chosen by the developer.
The release positions GPT-Live-1 for real-time customer service and voice-agent work, where interruption handling and low latency matter as much as conversational depth. The banking benchmark pass rate, still below one in three even after the reported improvement, indicates how much remains unsolved in regulated, script-heavy voice support. OpenAI has not published independent evaluations of the model.