OpenAI has released GPT-Live-1 in the API as a front-end voice layer that can listen and speak simultaneously, handle interruptions and delegate deeper work to a separate model.
RELATED BUYER PROFILEReview GPT-Live-1 pricing, fit and test-pending statusPricing · access · strengths · limitations →CONTINUE THE DECISIONBuild an AI stack for a customer support teamEvidence · workflow · next action →CONTINUE THE DECISIONDesign a measurable voice-agent pilotEvidence · workflow · next action →CONTINUE THE DECISIONCalculate complete cost per resolved callEvidence · workflow · next action →What you need to know
- GPT-Live-1 is available in the API at $0.05 per minute for the front-end voice layer
- It handles incoming and outgoing audio in one full-duplex model and supports interruption, turn detection and transcripts
- Reasoning and tool calls can be delegated to a separate OpenAI or third-party backend model, adding cost and another reliability boundary
The voice-agent buying brief
What changed for voice-agent builders
GPT-Live-1 is now available through the OpenAI API. Unlike a chained speech-to-text, language-model and text-to-speech pipeline, it processes incoming and outgoing audio together. OpenAI positions this full-duplex design for more natural pauses, interruptions, acknowledgements, background noise and longer conversations.
The voice layer can delegate the difficult work
Developers choose the model, tools and agent harness behind the conversation. GPT-Live-1 can hand deeper reasoning or actions to GPT-6 Astra, another OpenAI model or a third-party model while the voice interaction continues. This flexibility is useful, but it means the customer experience depends on both the real-time layer and the backend workflow.
What the $0.05 price does and does not include
OpenAI lists GPT-Live-1 at $0.05 per minute for the front-end voice layer. Backend model inference, tools, telephony, infrastructure, storage and human escalation are separate cost drivers. A ten-minute call is therefore not automatically a $0.50 resolved interaction; buyers should calculate cost per successfully completed outcome after transfers, retries and review.
Vendor evidence needs a same-call test
OpenAI reports a 30-point gain over GPT-Realtime-2.1 on its Full Duplex Bench and cites Speak finding almost 80% fewer interruptions than a previous turn-based system. These are OpenAI and customer-reported results, not independent HubAI measurements. Test the same accents, noise, pauses, corrections, tool actions and escalation cases your users will create.
HubAI buyer verdict
GPT-Live-1 is a strong candidate where conversational flow materially affects conversion, support or accessibility. It is not automatically the best choice for simple scripted calls. Compare it with a well-tuned cascade on completed-task rate, interruption errors, user correction, latency, escalation and total resolved-call cost. HubAI is withholding a score until that controlled test is complete.
HUBAI VIEWThe headline rate covers only the voice layer. Buyers should compare complete resolved-call cost, interruption quality, backend inference and human escalation—not price per minute alone.
Buyer decision signal: New API model · real-time voice
What to verify next
1Measure complete resolved-call cost, not only voice minutes
2Test accents, languages, background noise, pauses and mid-sentence interruption
3Separate front-end voice failures from backend reasoning and tool failures
4Define consent, recording, transcript retention and escalation rules
5Confirm telephony, regional and data-handling requirements for the intended market
6Compare a full-duplex build with the current STT-model-TTS cascade
7Keep consequential actions behind explicit confirmation and auditable logs
Read the evidence
Capabilities, availability and prices can change. HubAI keeps analysis separate from the underlying official material.
01OpenAI: Build natural voice experiences with GPT-Live-1Open source ↗02OpenAI Developers: GPT-Live-1Open source ↗03OpenAI API pricingOpen source ↗
Products
Products
Code AI
Image AI