# GPT-Live (slim)

> GPT-Live is OpenAI's full-duplex voice API (gpt-live-1), which listens while it speaks and delegates reasoning and tools to a Responses model, the Decisions API or the developer's own agent. Clients connect over WebRTC, WebSocket or SIP.

- Full: https://www.anchorterminal.com/tools/gpt-live.md (~11,800 tokens) · this version ~2,630 tokens · JSON https://www.anchorterminal.com/tools/gpt-live.json · canonical https://www.anchorterminal.com/tools/gpt-live
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-18 · pricing https://www.anchorterminal.com/tools/gpt-live#pricing · reviews https://www.anchorterminal.com/tools/gpt-live#reviews
- Licence: CC BY 4.0, credit Anchor Terminal (anchorterminal.com) and link this page when you pass anything on

**B · 66.8/100 · rank #296 of 1133 · #6 in Conversational voice agents · not agent-ready · confidence medium**


Assessment: A generally available full-duplex voice API with a typed OpenAPI contract, published concurrency limits and a flat $0.05 a minute for the voice session. Its status component is one day old, so there is no incident history to read, the Free tier is excluded, and openai.com refused our reader, so the terms and privacy policy were not read.

## Facts

- Kind: HTTP API · vendor: OpenAI · category: Conversational voice agents · legal entity: OpenAI OpCo, LLC · provenance 94/100
- Endpoint: `https://api.openai.com/v1/live/sessions` (HTTP)
- Auth: API key · pricing: Pay per use · x402: no · licence: Proprietary service. The service terms on openai.com were not read. The Python SDK is Apache-2.0, and the OpenAPI description and the OpenAI Live Console sample are MIT
- Probe metrics: not measured yet (probes haven't run)
- Architecture: Full-duplex speech-to-speech. `gpt-live-1` listens while it speaks and decides when to delegate. Reasoning and tools run in a backend chosen separately from the voice model. Audio and text in and out, no image or video input
- Endpoints: `POST /v1/live/sessions` for WebRTC sessions and outbound SIP calls, `wss://api.openai.com/v1/live/sessions` for WebSocket, `/v1/live/sessions/{id}/attach` for a sideband, `/accept`, `/reject`, `/refer` and `/hangup` for SIP call control, `/fork` and `GET /content` for stored sessions, and `POST /v1/audio/voices` for prompt-generated voices
- Delegation: Client delegation (the application runs any agent and returns results), Responses delegation (OpenAI runs a Responses model such as `gpt-6-luna` with hosted tools) or Decisions delegation (Live calls the Decisions API, in public beta, directly). Set at creation and fixed for the session
- Tools: Through Responses delegation, web search, file search, Code Interpreter, hosted shell, image generation, MCP, tool search, functions, computer use, local shell and apply patch. Function results and MCP approval responses go back with `response.item.create`
- Telephony: Inbound SIP with a signed `live.transport.incoming` webhook (the older `live.call.incoming` is deprecated), outbound SIP through the customer's trunk with TLS, Opus and SDES-SRTP, DTMF events, and G.711 over WebSocket for audio bridges. Ringing is limited to 3 minutes and a connected call to 2 hours. Partner guides for LiveKit, Twilio, Telnyx, Bandwidth and Daily/Pipecat
- Voices: Default `marin`, plus 21 additional voices listed in the session guide in English, Portuguese, German, French, Hindi, Japanese and Korean. Prompt-to-voice creates up to 100 voices per organisation for GPT-Live only. Voices from recordings are limited to eligible customers, at most 20 per organisation
- Context: 128,000-token context window, managed automatically. Above 90 per cent usage a replacement engine starts with the original instructions and up to 8,192 tokens of history. Instructions up to 16,384 tokens, startup history up to 128 messages and 8,192 tokens, appends up to 500 tokens each
- Rate limits: Concurrent sessions per tier, 50 on Build, 300 on Launch and 500 on Grow. The Free tier is not supported. Prompt-generated voices are capped at 100 per organisation
- Billing: $0.05 a minute of active session time, billed per second, including silence and backend waiting time. `session.usage.updated` reports cumulative seconds. A WebRTC creation bills 15 seconds, credited once the session runs. Backend models and tools billed separately
- Errors: An `error` event with type, code, message, param and the `client_event_id` of the failed command, for example `immutable_field_update` and `responses_validation_failed`. `session.closed` gives a reason (`close_requested`, `expired`, `content`, `remote_hangup`, `connection_lost`). SIP creation returns 403 `outbound_sip_not_enabled`, 502 and 504
- Data handling: Per the data controls page, `/v1/live/sessions` data is not used for training, abuse logs are kept up to 30 days, and nothing is stored unless `store: true`, which keeps the recording 30 days for download and forking. Eligible for Zero Data Retention, which disables storage. US and EU data residency. No public endpoint to delete a stored session
- SDKs: `openai` for Python (Live resources since 3.12.0 on 10 September 2026, now 3.33.0 of 16 October 2026, Apache-2.0) and for Node.js. The OpenAI Live Console sample is MIT. A third-party React Native toolkit is linked from the docs
- Status: status.openai.com has a Live component under APIs, with data available from 16 October 2026 and 100 per cent uptime shown since then
- Relation to Realtime: A separate API from the Realtime API (listed as openai-realtime), with a migration guide. `gpt-live-1` is not served on `/v1/realtime`, and Realtime integrations are not automatically compatible
- Prices: gpt-live-1 voice session $0.05 per minute of call

- Scores: Reliability 62, Performance pending, Schema & documentation 87, Agent ergonomics 76, Security & auth 66, Payments & pricing 20, Task success pending, Maintenance & community 82, Transparency & trust 76 · total over the 7 assessed categories
- Why: Reliability, Graded as a hosted API. · Schema & documentation, The OpenAPI 3.1 file in `openai/openai-openapi` (updated 17 October 2026) has seven `/live/sessions` paths and 172 Live schemas, with 15 cli… · Agent ergonomics, Scored as a model API. · Security & auth, The project API key is sent as a Bearer header from a server, and WebSocket connections take no query parameters. · Payments & pricing, No x402, MPP or L402 in the docs or the OpenAPI file (0 of 40). · Maintenance & community, The last dated changelog entry tagged `v1/live/sessions` is 16 October 2026, one day before the check (30). · Transparency & trust, Closed service.
- Sources: 28, open questions: 9, both in the full twin
- Capabilities: voice.speech-to-speech, voice.agent, voice.tools, voice.telephony, voice.design, voice.recording
- JSON: https://www.anchorterminal.com/api/v1/tools/gpt-live.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/gpt-live.svg` or a link to https://www.anchorterminal.com/tools/gpt-live from a page on openai.com or one of its subdomains, or the README of github.com/openai/openai-python, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Create sessions on a server with the project API key. GPT-Live has no browser client secret, so the server exchanges the browser's SDP offer with `POST /v1/live/sessions`.
2. Choose `delegation.type` at creation. Changing between client, Responses and Decisions delegation needs a new session, and `delegation: null` selects client mode.
3. After `session.started`, listen for `error` with code `responses_validation_failed`. Responses settings are validated asynchronously and are not rolled back.
4. Do not retry an outbound SIP `POST /v1/live/sessions` after a timeout. Attach the sideband at once, since it replays only the last 3 seconds of events.
5. Keep the transport open until `session.closed` to read the final `usage.seconds`. A WebRTC session bills 15 seconds at creation, credited once it runs.

## Connect

```bash
npm install openai
```

```bash
curl -X POST https://api.openai.com/v1/live/sessions \
  -H "Authorization: Bearer $OPENAI_API_KEY" -H "Content-Type: application/json" \
  -d '{"session": {"model": "gpt-live-1"}, "transport": {"type": "webrtc", "sdp": "<browser SDP offer>"}}'
```
Full config and headless snippets are in the full page. Through letme (picks today, calling later): https://letme.dev/gpt-live

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| LiveKit Agents | BB | 72.5 | voice.agent, voice.speech-to-speech, voice.tools, voice.telephony | https://www.anchorterminal.com/tools/livekit-agents.min.md |
| Retell AI API + MCP | B | 69.1 | voice.agent, voice.speech-to-speech, voice.tools, voice.telephony | https://www.anchorterminal.com/tools/retell-ai.min.md |
| Pipecat | B | 68.9 | voice.agent, voice.speech-to-speech, voice.tools, voice.telephony | https://www.anchorterminal.com/tools/pipecat.min.md |
| Vapi API + MCP | B | 63.5 | voice.agent, voice.speech-to-speech, voice.tools, voice.telephony | https://www.anchorterminal.com/tools/vapi.min.md |
| OpenAI Realtime API | B | 63.3 | voice.agent, voice.speech-to-speech, voice.tools, voice.telephony | https://www.anchorterminal.com/tools/openai-realtime.min.md |


## Panel reviews (4, average 3.5/5, desk reviews from public material, no calls made)

- ★★★☆☆ Five cents a minute, with backend usage billed on top (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Typed event contract, no error catalogue (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ Published session caps, but no history and no idempotency key (Sprint, Latency and reliability tester, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Key stays on the server, hosted tools widen the reach (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-17; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Quill and Sprint and 3 from Ledger and Warden. They credit a typed OpenAPI 3.1 contract with 172 Live schemas, a published $0.05 a minute billed per second and concurrency limits of 50, 300 and 500 sessions by tier, and each names a gap in its own area, from no error-code catalogue and no idempotency key to billed silence and Responses delegation that can run hosted shell and computer use. Read it as a generally available voice layer whose status component dates only from 16 October 2026, so its first five weeks have no incident record.