Move from GPT-Live-1 to Fullplex

A Fullplex live session uses the GPT-Live-1 session protocol. A client that works with GPT-Live-1 works with Fullplex when you change three values: the address, the API key, and the model. This guide gives those changes, what stays the same, and each difference that your client must handle.

Before you start

  • Complete Get access to get a Resia API key. The key needs editor or owner scope.
  • Keep the API key on your server. Never put it in a browser or a mobile app. Your web or mobile app sends its audio through your server, and your server opens the session.

Change three values

Value GPT-Live-1 Fullplex
Address host api.openai.com api.resia.ai
API key Your OpenAI API key Your Resia API key
Model gpt-live-1 resia-fullplex-1

The Fullplex address is wss://api.resia.ai/v1/live/sessions. Send the key in the Authorization: Bearer header, as before.

If your client sets a voice, change it too. Fullplex does not accept GPT-Live-1 voice names such as marin. See Voices.

Before and after

This server-side Node.js example uses the ws package. The first block opens a GPT-Live-1 session:

import WebSocket from "ws";

const ws = new WebSocket("wss://api.openai.com/v1/live/sessions", {
  headers: { Authorization: `Bearer ${process.env.OPENAI_API_KEY}` },
});

ws.on("open", () => {
  ws.send(JSON.stringify({
    type: "session.start",
    session: {
      model: "gpt-live-1",
      instructions: "Help the caller.",
      audio: {
        format: { type: "audio/pcm", rate: 24000 },
        output: { voice: "marin" },
      },
    },
  }));
});

The second block opens the same session on Fullplex. Only the address, the key, the model, and the voice change:

import WebSocket from "ws";

const ws = new WebSocket("wss://api.resia.ai/v1/live/sessions", {
  headers: { Authorization: `Bearer ${process.env.RESIA_API_KEY}` },
});

ws.on("open", () => {
  ws.send(JSON.stringify({
    type: "session.start",
    session: {
      model: "resia-fullplex-1",
      instructions: "Help the caller.",
      audio: { format: { type: "audio/pcm", rate: 24000 } },
    },
  }));
});

Without audio.output.voice, the session uses the default voice, Cartesia Skylar.

What stays the same

  • One connection is one session. Your client sends session.start as the first event and gets session.started.
  • Start fields. Fullplex accepts model, instructions, input, audio.format, audio.output.voice, delegation, and store.
  • Audio formats. Input and output use one format: audio/pcm at 24000 or 16000 Hz, or audio/pcmu or audio/pcma at 8000 Hz.
  • Client events. session.input_audio.append, session.input_audio.mute, session.input_audio.unmute, session.commentary.append, session.instructions.append, session.thinking.append, response.item.create, response.create, session.update, and session.close keep their names and fields.
  • Server events. session.started, session.updated, session.output_audio.delta, session.input_transcript.delta, session.output_transcript.delta, session.delegation.created, response.event, session.usage.updated, session.closed, and error keep their names.
  • Delegation types. A client delegation ({"type": "client"}) and a Responses delegation ({"type": "responses", "responses": {...}}) work as before. The delegation type cannot change after the session starts.

Connect a live session gives every start field and its limits. Session events gives every event and error.

Differences

Voices

Fullplex speaks with Cartesia and Fish voices. Set session.audio.output.voice as cartesia:<voice ID> or fish:<voice ID>, or leave it out for Cartesia Skylar. A GPT-Live-1 voice name gets an error with code invalid_value. GET /v1/live/available-voices lists the voices. See Session records and voices.

Fish voices are not available to an organization that has a business associate agreement (BAA) or zero data retention (ZDR) with Resia.

Start fields that Fullplex does not serve

Fullplex refuses these GPT-Live-1 settings. The server sends an error that names the field in error.param, then closes the connection.

Setting Fullplex
client.data_channel Not served. A Fullplex session runs on one WebSocket.
safety_identifier Not served.
triggers Not served.
delegation.type: "decisions" and delegation.decisions Not served. Use a client or a Responses delegation.
delegation.acknowledgement_delay_ms Not served.

Responses delegation settings

delegation.responses accepts model, instructions, tools, tool_choice, parallel_tool_calls, max_output_tokens, reasoning, and text. The backend model must be a model that is available to your organization.

Fullplex refuses service_tier, file search, code execution, provider file references, stored prompts, conversations, and previous_response_id. Function tools execute in your application, as before. See Delegation.

Delegation mode

Fullplex adds session.delegation.mode. It sets how often the session delegates:

Mode Behavior
auto, the default The session selects when to delegate.
always The session delegates its replies to completed turns of the person.
off The session does not delegate.

A client that leaves out mode gets auto. The mode cannot change after the session starts.

Events that Fullplex does not serve

Fullplex has no Decisions and no audio fixtures. These client events get an error with code unknown_event, and the session continues:

  • decision.create
  • decision.item.create
  • decision.item.delete
  • session.output_audio.fixture.play

Fullplex never sends fixture.playing or fixture.played.

Events that Fullplex adds

  • session.output_audio.cleared. Fullplex sends this event when it stops a reply while some of the reply's audio has not played yet. Stop the playback of the current reply and drop its queued audio. Audio after the event belongs to the next reply.
  • session.commentary.appended, session.instructions.appended, and session.thinking.appended. These events confirm each context append.

Output transcript

session.output_transcript.delta follows the audio that your client plays, not the audio that the server sends. The text of a reply part arrives when its audio starts to play, so the text can arrive after the audio events that carry it. When Fullplex clears a reply, it sends no text for the cleared audio.

Admission errors

Fullplex checks your Resia API key, your balance, your key scope, and the number of sessions that your organization has open. A refusal comes with an HTTP status and an error.code: 401 invalid_api_key, 402 insufficient_funds, 403 insufficient_scope, 429 session_limit, or 503 server_error. See Admission errors.

An organization can have five sessions at the same time by default. GET /v1/rate-limits reports the limit. See Rate limits.

Language

Fullplex speaks English only.

Recordings and session records

With store: true, Resia saves a stereo recording of the session. Download it with GET /v1/live/sessions/{session_id}/content after the session closes. GET /v1/live/session-history lists your sessions. See Session records and voices.

Price

A Fullplex live session costs 2.5 cents per minute, billed by the second. A Responses delegation adds the cost of its backend model tokens and of its hosted tools, such as web search. Resia bills the session to your organization's balance. GET /v1/live/sessions/{session_id}/charge returns the charge of one session. See Fullplex prices and Billing.

Check your client

  1. Open a session with the three new values.
  2. Send audio, and confirm that session.output_audio.delta and both transcripts arrive.
  3. If your client uses a voice, confirm that session.started shows your Cartesia or Fish voice.
  4. Remove any use of the events and settings in Differences that Fullplex does not serve.
  5. Handle session.output_audio.cleared in your audio player.
  6. Read the session charge after the session closes.