metir
metir
Docs
Download on App StoreGet it on Google PlayLoginSign Up
Back to Blog
OpenAI
Voice AI
AI Agents
API Pricing

GPT-Live-1 API: How to Test Costs and Cancellations

GPT-Live-1 API brings full-duplex voice to developers. Understand the September launch, separate reasoning costs, and what to test before deploying an agent.

Metir AI TeamSeptember 14, 20264 min read
GPT-Live-1 API: How to Test Costs and Cancellations

The GPT-Live-1 API arrived on September 10, 2026, bringing OpenAI's full-duplex voice model to developers. It can listen while speaking and pass deeper work to a separate model. The practical question is whether that smoother conversation also produces a correct, affordable outcome. The headline price covers only part of the system. OpenAI's announcement establishes what launched.

Our GPT-Live-1 launch analysis covers the architecture. This follow-up focuses on pilot costs, corrections, and cancellation behavior.

What changes with the GPT-Live-1 API

OpenAI separates conversation from task execution. GPT-Live handles spoken interaction; a backend handles reasoning, tools, and business instructions. Developers can use managed Responses delegation or connect their own model, agent framework, or service through client delegation. The getting-started guide explains this division.

Consider a caller checking an order who adds, “Actually, use my other address,” while a lookup runs. A useful assistant should hear the correction, retain it, and apply it to the right task. Keeping audio flowing solves the conversational part. Making the eventual action correct still requires application logic.

That distinction makes this release relevant to teams with existing agents: they can evaluate a new spoken interface while retaining their backend workflow. It also makes migration a design exercise, because the conversation and the work now progress independently.

Five cents is the voice layer

OpenAI's model page lists voice sessions at $0.05 per minute, billed per second, with backend model and tool usage charged separately. At that rate, the following are calculated voice-only examples, not complete service quotes:

$0.5010-minute sessionVoice layer only
$360-minute sessionVoice layer only
$50010,000 minutesBefore backend and tools
Voice session
Minutes × $0.05
+
Backend models
Separate usage charges
+
Tools
Separate usage charges
Core API cost components. Add any telephony, hosting, or other service charges to your budget; this is not an all-inclusive price.

A budget should therefore separate session duration, delegated reasoning, and additional services. The same ten-minute call could involve one simple lookup or repeated research and several actions. Its voice charge can match while its total cost differs.

For a pilot, record cost per successfully completed task alongside cost per minute. A cheaper conversation that requires a second call may be a worse outcome. Equally, a more expensive reasoning model may be unnecessary for a straightforward status check. These are evaluation questions to answer with representative calls, rather than assumptions to put into a savings forecast.

Sam Altman seated on a TED stage wearing a headset microphone
Sam Altman at TED, in an archival photograph. This does not depict the September 2026 API launch. Photo: Steve Jurvetson, CC BY 2.0.

Interrupting speech is not cancelling work

OpenAI's integration guide explicitly says that interrupting speech does not automatically cancel backend work. The application owns permissions, confirmations, and durable task state.

For a booking assistant, “wait” needs a defined meaning. Has the reservation merely been proposed, is submission underway, or has it already succeeded? A natural spoken acknowledgement should reflect that actual state. Otherwise, a caller may reasonably believe an action stopped when it continued.

“

Test what the assistant did, alongside what it said.

metir analysis

A focused pilot should include these scenarios:

  • Corrections during work: change a date or address during a lookup and inspect the eventual result.
  • Cancellation near submission: interrupt before and after an action starts, then verify the recorded outcome.
  • Conversation under pressure: try pauses, overlapping speech, background noise, and the accents your audience uses.
  • Unfinished sessions: disconnect during a task and check whether its status remains understandable.

A better way to judge voice agents

OpenAI reports that Speak's early evaluations reduced interruptions during thinking pauses by almost 80% compared with its previous turn-based systems. That is a customer-specific finding in the launch post, not a universal improvement guarantee.

The useful takeaway is to measure conversational quality and task quality separately. Start with a bounded workflow, compare successful completion and total costs, and inspect the moments when callers change their minds. Readers following the broader shift can find related analysis on the metir blog. GPT-Live-1 makes the voice layer more flexible; the pilot should establish whether the complete experience earns that flexibility.

Sources:

  • OpenAI: GPT-Live-1 API launch, September 10, 2026
  • OpenAI: GPT-Live-1 model and pricing
  • OpenAI: Getting started with GPT-Live

Image credits

  • Sam Altman at TED: Steve Jurvetson via Wikimedia Commons, CC BY 2.0. Archival photograph, resized by Wikimedia; used for the hero and in-body figure.

Ready to experience AI that adapts to you?

metir brings together the world's best AI models in one seamless experience. Start for free today.

Get Started Free
metir

Agentic Operating System for Professionals buried in meetings, emails and docs.

© 2026 metir. All rights reserved.

Product

  • Features
  • Pricing
  • Research
  • Docs
  • Blog
  • Enterprise

Company

  • Docs
  • Support
  • Careers

Legal

  • Terms of Service
  • Privacy Policy

Personalisation is powerful. Privacy is non-negotiable.

Status: All systems operational