GPT-Live-1.1 Lands in ChatGPT; the API Stays on GPT-Live-1

OpenAI shipped GPT-Live-1.1 to ChatGPT voice on October 15, but the GPT-Live-1.1 API is not available yet. The developer docs still list only gpt-live-1. If you are scoping a voice feature, price and plan against that model, and keep the model id in config so the upgrade is a one-line change.

What OpenAI shipped on October 15

OpenAI's community announcement says GPT-Live-1.1 is "rolling out today to all users" in ChatGPT, Codex and dots. The listed changes:

• 13 new voices.
• Better performance in French, German, Hindi, Mexican Spanish, Korean, Japanese and other non-English languages.
• More natural delegation and better steerability.
• Custom voices from a text description or an approved recording, added in an October 16 update.

AI Weekly also reports a smaller variant called GPT-Live-1.1 Quip. The community post doesn't mention Quip, so treat that as secondary reporting.

GPT-Live-1.1 API status: coming soon, no date

The same post says "Coming soon to the API," and an OpenAI reply in the thread repeats it. There is no date, no price and no model id. AI Weekly says API access is "still pending."

The gpt-live-1 model page lists gpt-live-1 as the only snapshot and never mentions 1.1. Any 1.1 API price you see quoted is a guess.

Users in the thread also say some voices (Sol, Spruce, Arbor) sound different or worse after the update, mostly on Android and in Standard mode. I found no official response. That is the consumer app, not the API, but it is a good reason to schedule a listening test before you switch.

What a gpt-live-1 session costs

From the model page:

• Voice: $0.05 per minute, billed per second, not rounded up to the next minute.
• Backend: model and tool usage is billed separately, at normal pricing for whatever model and tools you configure.
• Concurrent sessions: Build 50, Launch 300, Grow 500. The Free tier is unsupported.
• Endpoint: Live (v1/live/sessions) only. Realtime, Responses and Chat Completions are listed as unsupported for this model.
• Features: streaming and function calling work. Structured outputs and fine-tuning do not.

VOICE_RATE_PER_MIN = 0.05  # gpt-live-1, billed per second; recheck the model page

def voice_cost(seconds: int) -> float:
    return seconds / 60 * VOICE_RATE_PER_MIN

print(round(voice_cost(277), 4))  # a 4m37s call -> 0.2308

This covers the voice line only. A thousand six-minute calls a month is 6,000 minutes, or $300 of voice time, before a single backend token. Budget the backend as its own line item.

Keep the model id configurable

The Create session reference puts the model in the session config, not in a URL query parameter. Voice and frontend instructions are fixed at startup, and so is the delegation mode: to change any of them, you start a new session. That makes the model id a natural environment variable.

import os
from openai import OpenAI

client = OpenAI()  # reads OPENAI_API_KEY

LIVE_MODEL = os.environ.get("LIVE_MODEL", "gpt-live-1")
BACKEND_MODEL = os.environ.get("LIVE_BACKEND_MODEL", "gpt-6-luna")


def start_live_session(browser_sdp: str):
    # Voice, instructions and delegation mode are fixed once the session starts.
    return client.live.create(
        session={
            "model": LIVE_MODEL,
            "instructions": "<frontend voice and conversation instructions>",
            "delegation": {
                "type": "responses",
                "responses": {
                    "model": BACKEND_MODEL,
                    "instructions": "<backend prompt with business rules>",
                },
            },
        },
        transport={"type": "webrtc", "sdp": browser_sdp},
    )

The response carries a session id and an SDP answer. Apply the answer in the browser, then wait for session.started on the data channel before sending commands. The backend model name comes from OpenAI's delegation guide, which suggests it as a starting point. Swap in your own.

Monday morning checklist

1. Price the pilot on gpt-live-1 and mark 1.1 as unpriced and undated in the plan.
2. Move the voice model id and the backend model id into config, as above.
3. Compare your peak concurrent calls with your tier's cap. Build allows 50.
4. Run a few real conversations and read the usage to size the backend cost.
5. If your prototype is on Realtime, plan a migration. Live is a separate endpoint, so it is not a model id swap.
6. Don't promise stakeholders the new voices. The docs I read don't say which of them the API will offer. Plan a listening test in your target languages for when 1.1 arrives.

Planning against what the docs list today, and treating 1.1 as a config change you haven't been handed yet, keeps a voice launch from slipping on a model with no API date.