Skip to content
AI Primer
release

OpenAI releases GPT-Live-1.1 through its API with $0.02 standby calls

OpenAI released its ChatGPT voice model to API developers with improved multilingual accents and instruction following. Standby mode keeps WebRTC calls connected for $0.02 per minute instead of $0.05.

5 min read
OpenAI releases GPT-Live-1.1 through its API with $0.02 standby calls
OpenAI releases GPT-Live-1.1 through its API with $0.02 standby calls

TL;DR

Muting the microphone still incurs active-session charges. The session guide warns that the first response after standby can take longer, while CellCog's test caught both voice models inventing reviewers for a renewal draft when they skipped the backend.

What shipped

  • Model: gpt-live-1.1 is a full-duplex model that listens while speaking, with audio and text input and output, according to the model documentation.
  • Speech recognition: OpenAI claims better handling of precise details, including letters and numbers, in its release announcement.
  • Multilingual voice: The model documentation claims more accurate transcription and more consistent accents and voice across languages than GPT-Live-1.
  • Instructions: Better instruction following and more reliable task completion are among OpenAI's stated improvements.
  • Active pricing: GPT-Live-1 → GPT-Live-1.1 stays at $0.05 → $0.05 per minute, 0% change, in the pricing table.

Standby

OpenAI's session guide describes three stages for a standby-enabled WebRTC session:

  1. Configuration: Speech-based wake-up uses session.wake_policy.type: "speech", with an idle_timeout_s value from 5 to 60 seconds.
  2. Pause: After the idle timeout, session.paused signals that inference has stopped. The WebRTC connection, sideband connection and session ID remain intact, while the voice rate changes from $0.05 to $0.02 per minute.
  3. Resume: Speech wakes inference, restores conversation history and produces session.resumed when inference is ready.

Without idle_timeout_s, inference stays active after the first wake until the session ends. The guide also documents an optional vad_threshold from 0 to 1, defaulting to 0.5; higher values require stronger evidence of speech at initial wake-up and after each pause.

Customer-reported benchmarks

OpenAI's model card provides no numerical benchmark scores for its transcription or multilingual-voice claims.

CellCog founder and CEO Nitish Garg reported a 680-call comparison using the company's production meeting configuration, changing only the model name. The first 560 calls alternated models across five scripted scenes:

  • Two direct questions addressed to the AI employee.
  • Two people talking without addressing it.
  • A direct question after other people talked.
  • A greeting and microphone check.
  • A silent room.

Two measures favored GPT-Live-1.1:

  • Median time to first sound: 1.3 → 1.2 seconds, -0.1 seconds, in CellCog's timing results.
  • Admissions of uncertainty among answers given without the backend: 0/26 → 14/67, or 0% → 20.9%, +20.9 percentage points, by CellCog's manual count.

The second metric is conditional: the models produced different numbers of answers without consulting the backend.

Startup delegation regressions

OpenAI promised more reliable task completion, while CellCog measured three regressions in its existing configuration. These comparisons run from GPT-Live-1 to GPT-Live-1.1:

A further 120 calls narrowed the handoff finding to startup timing. In CellCog's follow-up:

  • First question at 0.5 seconds: Handoffs fell from 45/60 to 28/60, or 75% → 46.7%, -28.3 percentage points.
  • First question at ten seconds: Both models handed off 48/60 questions, or 80% → 80%, 0 percentage points.

The silence regression also had a narrower shape: GPT-Live-1.1 made more one- or two-word interjections, but spoke three words or more in fewer calls, according to CellCog's breakdown.

The tests used synthetic English speakers and short calls, with prompts tuned for GPT-Live-1. They did not evaluate multilingual accent quality, long conversations or standby costs.

Backend delegation and billing

GPT-Live can hand reasoning and tool work to a backend model chosen by the developer, according to OpenAI's announcement. In Vercel's client-delegation example, the application controls the workflow:

  1. Session instructions define when the voice model asks for help.
  2. The application receives the delegation request and calls its selected text model.
  3. The result returns through the commentary channel for spoken output.

Permissions, confirmations and cancellation remain application responsibilities in that example.

Voice sessions are billed per second, with backend models and tools charged separately. OpenAI's cost documentation counts these periods as active time:

  • Speech.
  • Silence before the idle timeout expires.
  • Waiting for backend work while inference is running.

The documented standby discount applies to WebRTC sessions while inference is paused.

Live endpoint limits

The model card lists two integration constraints:

  • Endpoint: v1/live/sessions is supported. Chat Completions, Responses and v1/realtime are unsupported as direct endpoints for this model.
  • Concurrency: GPT-Live-1.1 and GPT-Live-1 have separate concurrent-session limits. Each model's quota applies across inference regions, with account-specific limits.

AI Gateway and the Live playground

Vercel kept the same Live endpoint and client-delegation interface, so existing GPT-Live integrations require only the new model ID, according to its launch notes. Its AI SDK example selects the Live API explicitly:

OpenAI also exposes the model in the Platform's Live tab, as testingcatalog reported.

Further reading

Discussion across the web

Where this story is being discussed, in original context.

On X· 2 threads
TL;DR1 post
AI Gateway and the Live playground1 post
Share on X