GPT-Live-1.1 Guide: Multilingual Audio, Standby, and Plain-Text Handoff

Learn how OpenAI's GPT-Live-1.1 full-duplex voice model improves multilingual accents and transcription, adds $0.02/min standby, and simplifies delegation.

By ailog · Published Nov 20, 2026 · 7 min read · How we work
In short
  • GPT-Live-1.1 upgrades multilingual voice interactions with refined pronunciation, accent consistency, and higher transcription accuracy.
  • Standby mode for WebRTC sessions drops voice billing from $0.05 per active minute down to $0.02 per minute while inference pauses during silence.
  • Prompting transitions away from legacy <|SpawnThinking|> tokens to plain-language delegation rules in session instructions.
  • The model operates strictly over the v1/live/sessions endpoint and maintains separate concurrent-session rate limits from GPT-Live-1.

What Changed in the GPT-Live-1.1 Release

On 2026-11-19, OpenAI introduced gpt-live-1.1 across the v1/live/sessions endpoint. The release represents an iterative update to the full-duplex audio architecture, specifically targeting speech transcription fidelity, pronunciation accuracy, and accent consistency across non-English and multilingual dialogues. Like its predecessor, the model processes streaming audio directly and maintains full-duplex capabilities, meaning it can process incoming user speech simultaneously while producing audio output.

Alongside conversational quality upgrades, the release formalizes support for WebRTC session standby controls and a simplified delegation schema. Instead of relying on rigid, synthetic control tokens to route tasks to background workers, gpt-live-1.1 adopts natural-language delegation rules defined directly inside session instructions. Teams building real-time interactive voice agents can review foundational voice concepts in our real-time voice APIs guide.

Live Session Pricing and Standby Economics

Unlike text completion endpoints billed strictly on token throughput, sessions running on v1/live/sessions bill according to elapsed connection time. Standard active interaction carries a flat per-minute cost, while silent or paused periods on WebRTC can take advantage of the standby discount. Any external tools or secondary models invoked to process background logic bill independently according to their respective pricing structures.

Session State Price Unit Billing Details
Active $0.05 minute Standard full-duplex voice streaming and live inference
Standby $0.02 minute WebRTC sessions with inference paused during user inactivity

The standby rate of $0.02 per minute applies exclusively to WebRTC sessions when audio inference is suspended. As soon as the user speaks or the session resumes active conversational processing, billing shifts back to the active rate of $0.05 per minute. If your voice architecture offloads background reasoning steps to secondary models like gpt-5.6-luna ($0.2 per 1M input tokens, $0.02 cached, $1.2 output) or gpt-6.1-sol ($2 per 1M input tokens, $0.1 cached, $10 output), you can model combined pipeline expenses using our /tools/llm-api-cost-calculator/.

Plain-Language Delegation vs Legacy Tokens

A notable developer experience shift in gpt-live-1.1 is how developers configure backend task delegation. Under gpt-live-1, prompts required explicit instruction tags, specifically mandating the model to produce <|SpawnThinking|> when background assistance was needed or instructing it Do not generate <|SpawnThinking|> when handling conversational replies directly. In gpt-live-1.1, this token-based syntax is replaced with descriptive natural language under dedicated instruction headings.

Session instructions accept up to 16,384 tokens and should be organized into clear behavioral policies. The recommended structure establishes rules for conversational cadence, backchannels (audible cues such as ‘mm-hmm’ that indicate active listening without yielding the floor), interruption behavior, and delegation criteria:

You are a calm, friendly voice assistant for support.
Speak warmly and naturally, at an unhurried pace.

Backchannel policy: Use moderate backchannels. Acknowledge naturally without competing with the main response.

Interruption policy: Stop speaking when the user interrupts. Listen to what they say.

Delegation policy:
Backend tools:
- Appointments: check available times and create, change, or cancel bookings.

Delegate to the backend when:
- The user asks for availability or wants to create, change, or cancel a booking.
- A correction changes a booking task already in progress.
- The answer needs careful reasoning beyond a simple reply.

Do not delegate when:
- The user greets you or asks you to repeat a result already provided.
- You cannot tell what they are asking for without a brief clarification.

Delegate to the backend before giving an answer that depends on backend work.
Do not guess the result while waiting.

Spoken promises such as ‘I am checking that’ do not automatically initiate backend tasks. The model must evaluate the plain-language boundaries specified under the delegation policy, trigger the handoff, and await results before reporting task completion to the caller.

Supported Voices and Multilingual Accent Management

To address pronunciation drift and inconsistent accents across international deployments, gpt-live-1.1 improves voice stability across diverse languages. Audio output configuration accepts both natural human captures and generated voices across various regional accents. You specify the voice at session creation using audio.output.voice, which defaults to marin if omitted.

Voice API Name Language Target Regional Influence Presentation Voice Source
quartz English Australian Feminine Generated
ripple English Australian Masculine Natural
vesper English British Masculine Natural
willow English Irish Feminine Natural
stone English Irish Masculine Natural
gleam English North American Feminine Natural
meridian English North American Masculine Natural
bossa Portuguese Brazilian Feminine Natural
tempo Portuguese Brazilian Masculine Natural
beacon English Filipino Masculine Generated
delta English Southern U.S. Feminine Generated
cinder English Southern U.S. Masculine Generated
nuri German — Feminine Natural
juni German — Feminine Natural
flitz German — Masculine Natural
brise French — Feminine Natural
sillage French — Masculine Natural
nira Hindi — Feminine Natural
harema Japanese — Feminine Natural
shida Japanese — Masculine Natural
noeul Korean — Feminine Natural

For configurations where built-in options do not match target demographics, teams can design custom voices via prompt descriptions or approved recordings. Details regarding voice customisation workflows can be reviewed in our guide on OpenAI prompt to voice and GPT-Live multilingual voices.

Session Management and Real-Time Event Architecture

Sessions must be established directly against the v1/live/sessions route. The model snapshot gpt-live-1.1 does not operate on legacy text completion routes, v1/chat/completions, v1/responses, standard v1/realtime, batch queues, or fine-tuning pipelines. Applications can seed initial conversation context using the input array at startup, which accepts up to 128 prior messages and 8,192 combined tokens.

from openai.types.live.session_config_param import SessionConfigParam

session: SessionConfigParam = {
    "model": "gpt-live-1.1",
    "input": [
        {
            "type": "message",
            "role": "user",
            "content": [
                {"type": "input_text", "text": "I need help with my recent order."}
            ],
        },
        {
            "type": "message",
            "role": "assistant",
            "content": [{"type": "output_text", "text": "What is the order number?"}],
        },
    ],
}

Once a session is established and emits session.started, live state is managed through three dedicated append events, each restricted to a maximum string payload of 500 tokens:

  • session.instructions.append: Appends trusted behavioral guidelines that alter the conversational or tone rules during runtime.
  • session.thinking.append: Injects background factual context silently without instructing the voice model to articulate it immediately to the caller.
  • session.commentary.append: Supplies updated operational details or progress notifications that the model is permitted to paraphrase aloud to the user.

During live dialogue, transcript events stream back to the client via session.input_transcript.delta for caller speech and session.output_transcript.delta for assistant audio. Each delta provides start_ms and end_ms millisecond offsets representing the relative timeline position from the start of the session rather than network transmission times.

Operational Constraints and Rate Limits

When adopting gpt-live-1.1, infrastructure teams must account for key architectural boundaries:

  • Independent Concurrency Pools: Rate limits for gpt-live-1.1 are calculated separately from gpt-live-1. Existing session concurrency allowances on the older model do not automatically roll over.
  • Immutable Session Parameters: The model, initial instructions, input context, and audio.output.voice cannot be modified mid-session. Changing these core parameters requires terminating the connection and creating a new session.
  • Protocol-Specific Audio Formats: WebSocket connections require declaring audio.format explicitly at connection initiation. Conversely, WebRTC handles format negotiation automatically during signaling, requiring audio.format to be omitted.
  • Context Timing Boundaries: An appended event’s acknowledgment occurs when the timeline reaches its estimated offset (start_ms to end_ms), but the model might generate audio before fully evaluating an incoming update. Critical permissions and transaction validations must remain anchored in your backend logic rather than conversational assumptions.

For teams integrating automated classification or pre-call policy routing alongside voice pipelines, compare relevant models across providers using our guide on choosing an LLM API model by cost tier and review broader API data retention and training policies.

Migration and Deployment Checklist

  • Session Initialization: Update the startup model property to gpt-live-1.1 in your session payload targeting v1/live/sessions.
  • Instruction Refactoring: Remove all legacy <|SpawnThinking|> and Do not generate <|SpawnThinking|> syntax from system prompts. Structure rules under explicit Backchannel policy, Interruption policy, and Delegation policy sections.
  • Standby Implementation: If connecting over WebRTC, implement inactivity detection to transition idle audio streams into standby mode, dropping the session billing rate from $0.05/min to $0.02/min.
  • Voice Selection: Verify chosen voices against target languages using regional options such as bossa for Brazilian Portuguese, nuri for German, or harema for Japanese.
  • Quota Verification: Inspect organization settings on the developer platform to ensure concurrent-session quotas are adequately provisioned for gpt-live-1.1 independently of older snapshots.

By aligning prompts to natural language delegation patterns and leveraging WebRTC standby controls, development teams can improve multilingual user experiences while managing operational costs.

Sources
  1. OpenAI API changelog — 2026-11-19
  2. GPT-Live-1.1
  3. Prompting GPT-Live
  4. Managing GPT-Live sessions

Related