GPT‑Live‑1 brings natural, two-way voice conversations to the API. The model can listen and speak at the same time; it can delegate tasks to the models and tools it is paired with for deeper reasoning and actions. Developers can tailor the voice experience to their users, workflows and goals.
The model’s ability to manage interruptions was viewed positively by Speak in early evaluations. GPT‑Live‑1 reduced interruptions by nearly 80% compared with previous sequential systems by giving students more time to think before the language instructor responded. Unlike traditional voice agents, which divide conversations into speech-to-text, reasoning and text-to-speech stages, GPT‑Live‑1 combines listening and speaking in a single model. It can manage responses, acknowledgments and interruptions in real time while continuing deeper processing in the background.
Developers can choose the models, tools and agent infrastructure; Luna can be used for high-volume scheduling or order updates, while Astra can be used for complex customer issues. This makes it possible to adjust speed, cost and reasoning level according to the task. GPT‑Live‑1 provides ASR transcripts and response text; it supports alphanumeric expressions and keyword prioritization. Although it is not a sequential model, turn detection is also built in.
In evaluations, GPT‑Live‑1 increased its Full Duplex Bench performance by 30 points against GPT‑Realtime‑2.1. It ranked first on Tau3 in mid-level reasoning with GPT‑6 Astra. The model offers a wider range of voice options across different accents, dialects and languages. It is available in the API today at a price of $0.05 per minute for the frontend voice layer. It can be integrated with Codex and OpenAI Presence; the sales team must be contacted for access to custom voices.
Why it matters
This approach is particularly relevant to education and customer service scenarios, where interactions are frequently interrupted, as it makes it possible to manage the flow of conversation in voice applications without dividing it into separate recognition, processing and speech synthesis steps. Its ability to handle listening, responding and background tasks within the same structure allows developers to build different components according to their requirements for speed, cost and reasoning rather than using a single type of agent. Results from early evaluations show that performance comparisons measure not only technical capacity but also the naturalness of conversation and wait times. The API’s ability to provide transcripts and response text makes it easier to connect voice interactions to existing workflows. However, the total cost of per-minute pricing at high volumes and the fact that access to custom voices depends on the sales process remain open points that teams moving to implementation need to assess separately.
Background
GPT is not a new name in the FikirPilot archive: we published 2 news articles mentioning it in the last 90 days; the most recent is dated September 13, 2026.
Term: agent
An artificial intelligence agent is software that calls tools and carries out multi-step tasks to achieve a goal instead of producing a single response.