Skip to content
· 3 min read

GPT-Live API still isn't a business receptionist

GPT-Live API shipped Sep 10, 2026 at $0.05/min. It is a full-duplex voice layer, not an $8k owned receptionist that books jobs and writes CRM notes.

A calm contractor front desk with an open paper appointment book, a handwritten job log, and a marker-drawn call-flow board, standing in for what the GPT-Live API still does not replace.
Article language

Showing original language

OpenAI put the GPT-Live API in developers’ hands on September 10, 2026. The voice layer is $0.05 per minute. It still does not answer your shop number, book the slot, or write the CRM note.

ChatGPT Voice already ran GPT-Live inside the app. OpenAI first described that model on July 8, 2026: full-duplex, listen and speak at the same time, backchannel “mhmm,” wait when you pause, and hand hard questions to a frontier model in the background. They said more than 150 million people talk to ChatGPT with Voice or Dictation each week. That is a consumer chat product. It is not a front desk.

What did the GPT-Live API actually ship?

GPT-Live-1 is a full-duplex speech model you can put in an app: talk over it, interrupt it, and keep the conversation going while a backend model thinks. You still have to own the phone number, the booking rules, and the system of record.

OpenAI’s own notes: GPT-Live-1 and GPT-Live-1 mini rolled into ChatGPT Voice first (paid tiers vs Free). The API version is the same idea for builders — one voice front-end, then delegation to whatever reasoning model and tools you wire. Extra voices, stronger instruction following, and telephony hooks. Transcripts and response text come out of the model. Backend usage bills separately.

I am repeating their launch copy, not running a bake-off. $0.05 a minute is the voice meter. If the backend looks up Jobber, HubSpot, or a calendar on every call, that meter stacks.

Does GPT-Live replace the business phone?

No. GPT-Live is the mouth and the ear. A receptionist is the job: answer the line, qualify, book or capture, write the note, and escalate the weird ones.

The leak I actually see: a plumber on a ladder, a salon owner in color, an attorney in court. ChatGPT on a personal phone does not take the inbound DID. It does not know your after-hours emergency fee. It does not write “burst pipe, 14th and Congress, caller Jane, call back in 12 minutes” into Housecall Pro.

If you want the SaaS-voice cost picture, I already compared ServiceTitan AI voice agent cost. Different wrapper, same question: what sits around the model.

The map I would still build:

  1. Trigger: inbound call or missed-call text on the business number.
  2. Action: the agent greets, asks two or three qualifying questions, books or holds.
  3. System of record: Jobber, Housecall Pro, GlossGenius, HubSpot, or a shared sheet — not a ChatGPT thread.
  4. Human: you (or a named tech) take price fights, emergencies that need a person, and anything the script does not cover.

That is AI employee replacement as tasks, not a ChatGPT subscription doing the owner’s job.

When should you ignore the demo and wait?

Skip a custom voice agent if you have no CRM, no booking rules written down, and no person who will take the 8pm escalation. A prettier conversation does not fix a messy book.

GPT-Live also is not Grok Bot. Grok Bot is a shared computer for desk work. It does not sit on the phone network. ChatGPT Voice is a chat you tap. Neither is the AI Receptionist I hand-deploy for $8,000 one time — you own the setup, no per-call fee to me.

The API change is real: builders can now rent a full-duplex mouth without stitching speech-to-text, a chat model, and text-to-speech. That is cheaper plumbing on my side. It does not change what you should buy. If the leak is missed calls, you still need the number, the calendar, and the CRM note. Send the short audit form; I reply with the replacement map within 24 hours.

Related operator notes

Keep reading

No-pressure first step

Not sure which one fits?
Get a free 20-min audit.

Bring one workflow you'd want automated. I'll tell you which deployment fits — and which doesn't — in twenty minutes. No pitch deck, no follow-up sequence. Useful even if you don't buy.

  • A real plan, not a sales call

    Which surface (Telegram, Discord, Slack, phone) fits your team, and which one doesn't.

  • Honest "don't buy this" if it applies

    If a $99/month SaaS solves it, I'll tell you which one and how.

  • A timeline + price range

    When I could deploy, what it'd cost, and what you'd own at the end.