AGENTIC COMMONSAI industry briefings

繁中EN

TopicAI Agents & WorkflowsPublished

All English articles

GPT-Live-1.1: Voice Sessions, Standby Pricing, Delegation

Summary: OpenAI shipped GPT-Live-1.1 on November 19, 2026, a full-duplex voice model that can listen while it speaks and hands tasks to a backend agent through a plain-language prompt. Compared with GPT-Live-1, the company says it transcribes speech more accurately and keeps accent and voice more consistent across languages. For builders of voice-first products — phone agents, support lines, multilingual assistants — the interesting part is the architecture: conversation and task execution now live in two layers you can prompt and swap independently.

What actually changed

The changelog entry describes the release as improving “pronunciation, accent consistency, and transcription in multilingual voice conversations.” These are OpenAI’s own claims; the documentation offers no third-party measurement, no list of which languages improved, and no benchmark. The model itself is narrow by design: it accepts audio and text in, returns audio and text out, and supports streaming and function calling. Nothing else — no Chat Completions, no Responses endpoint, no batch, no fine-tuning, no standalone audio endpoints. It runs only through v1/live/sessions.

One change from the same entry applies to both GPT-Live-1 and 1.1: WebRTC sessions can be configured for standby, pausing inference during inactivity and resuming when someone speaks. That is what unlocks the cheaper rate below.

How delegation works

GPT-Live handles the spoken conversation. A backend — a separate model or agent — does the reasoning, tool use, and task work, then returns results for GPT-Live to communicate. The getting-started guide gives the canonical example: a caller asks about an order and adds a detail while the backend checks status; GPT-Live keeps talking and explains the result when it arrives. If the caller interrupts, the backend work can continue — your application decides whether to finish or cancel it.

You pick the backend in one of two modes:

  • Responses delegation — OpenAI runs the backend model and passes conversation context back and forth. Your application still runs your own function tools.
  • Client delegation — you connect any model, agent harness, or service your application runs, and you control the backend’s context and execution.

The catch: you choose the mode when you create the session, and to change modes you start a new session. That is a per-session commitment, not a per-conversation toggle. Your prompt decides when GPT-Live delegates, in plain language.

What it costs and what it locks in

Voice sessions bill by duration: $0.05 per minute active and $0.02 per minute standby, with the standby rate applying to WebRTC sessions while inference is paused. Backend model calls and tool usage bill separately — so the per-minute number is a floor, not the cost of a session.

Two more operational facts from the model page. First, GPT-Live-1.1 and GPT-Live-1 have separate concurrent-session limits, and each model’s limit applies across all inference regions — upgrading doesn’t inherit your 1.0 capacity. Second, if you already run calls on LiveKit, Twilio, Telnyx, Bandwidth, or Daily/Pipecat, OpenAI documents partner integrations that connect your existing audio streams; WebRTC is the recommended connection for browser, mobile, and server, with WebSockets and telephony/SIP as alternatives.

The single-endpoint restriction is a lock-in worth naming. You cannot route GPT-Live-1.1 through your existing text pipeline, and leaving later means rebuilding around the Live session model. If you benchmark switching costs the way we compare voice agent platforms by cost per resolved call, add that rebuild to the ledger.

Our take: If you ship a spoken interface, spend an afternoon on the WebRTC quickstart — it is a cheap, reversible test, not a migration, and the bar for that kind of experiment is low. We would not treat the transcription and accent-consistency improvements as settled: they are the vendor’s own unmeasured words, so the test that matters is whether a real session in your languages completes the user’s task cleanly. Cost is the second gate. At $0.05/min active plus per-call backend billing, a busy multilingual session could run several times the per-minute rate once delegation fires often — we would instrument one day of real calls before committing. If your current chained or text-based pipeline already completes the task reliably, skip this; the architecture is elegant, but the finished answer is the only metric. If all-in session cost exceeds your current stack, or the multilingual gains don’t hold on your languages, that reverses the recommendation.

Sources

Original post
AGENTIC COMMONSOperated by PHLEGON LABS
ShareXEmail
Support us

Related reading

  1. ElevenLabs Cue: Hour of Audio Aligned in Half a Second

    ElevenLabs' Cue aligner adds character-level timestamps, handles any file length in one pass, and aligns an hour of audio in under a second.

    Voice AI

  2. Amazon Bee Frontier and Eclipse: Chat That Turns Into Shopping

    Amazon's Bee Frontier and Eclipse listen to daily conversations and turn them into reminders, summaries, and opt-in Amazon shopping suggestions.

    AI

  3. AI voice agent platforms: compare cost per resolved call

    Judge AI voice agent platforms by cost per resolved call: Firecrawl's breakdown turns a $0.05 platform minute into $0.13, $0.78 per six-minute call.

    Voice AI