GPT-Live-1 hits the API at $0.05 a minute
OpenAI shipped GPT-Live-1 to the API at $0.05/minute billed by the second, with backend model and tool calls priced separately. Run the math on your voice agent.
OpenAI put GPT-Live-1 in the API on September 10, priced at $0.05 per minute and billed by the second. OpenAI's pricing page is explicit about the part that will surprise you: backend model and tool usage is charged separately. The full-duplex voice model we covered in July is now something you can point at your phone line — and it finally has a number attached.
What actually happened
GPT-Live-1 is the API version of the full-duplex conversation model that shipped in ChatGPT. Full-duplex means it processes incoming audio while it is still speaking, so it can listen, pause, interrupt, or fire a tool call mid-sentence instead of waiting for a turn to end. That is the difference between a voice agent that sounds like a phone tree and one that sounds like a person who can be interrupted.
It runs on its own Live Sessions endpoint rather than through Chat Completions or Responses, so an existing realtime integration is a port, not a config change. It ships with a set of voices covering multiple accents and languages, and it emits transcripts alongside audio.
The pricing structure is the thing to internalize. $0.05/minute buys you the voice layer. Reasoning handed to a backend model, and any tools that model calls, bill at their own rates on top.
Why voice agent pricing is not $0.05 a minute
Do the arithmetic before someone else does it for you. A four-minute support call is 20 cents of voice. But a useful call — one that looks up an order, checks inventory, reschedules an appointment — is 20 cents of voice plus every backend token and tool invocation that made it useful. Depending on how chatty your orchestration is, the voice layer can be the smaller line item.
That is not a complaint. It is a design constraint, and it points at an obvious optimization: keep the voice model doing voice, and keep the reasoning narrow. Most bad voice agent economics come from routing the entire conversation through an expensive backend on every turn, when the majority of turns are "yes," "the blue one," and "can you repeat that."
The operator math is still favorable. A Phoenix answering service runs $1–2 per call. A four-minute AI call with real tool use lands well under that. The gap is not in the token price — it is in the 15% of calls the agent should hand to a human, and whether your build has a clean handoff or a dead end.
Price the whole call, not the minute. Then decide.
Key takeaways
- GPT-Live-1 is generally available in the API as of September 10, at $0.05/minute billed by the second with no rounding
- Backend model calls and tool usage bill separately — the voice layer is often not your largest cost
- Full-duplex: the model listens while speaking and can interrupt, pause, or call tools mid-response
- It uses a dedicated Live Sessions endpoint, not Chat Completions, Responses, or the Realtime endpoint — existing integrations need a port
- Keep reasoning narrow: most conversational turns don't need an expensive backend call
- Model the full call cost against your current answering service, and budget for the human handoff path
Run the numbers before you build the bot. Our ROI calculator models call volume, handle time, and escalation rate so you can see what a voice agent actually saves you — including the calls it should never take. Model your call volume, or tell us what your phone line costs you today.
Sources: OpenAI API pricing, OpenAI: GPT-Live-1 in the API.
- #openai
- #voice-ai
- #gpt-live
- #ai-pricing
- #phone-automation
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Zscaler Agentic SOC: the agent now pulls the trigger
Zscaler's Agentic SOC uses Anthropic and OpenAI models to triage, investigate, and contain threats automatically. The line that moved is autonomous remediation.
Read itPositron raises $875M for inference chips without HBM
Positron AI raised $875M at a $5B valuation for inference silicon built on commodity LPDDR5X instead of HBM. Why the memory choice decides your token price.
Read it