OpenAI's voice model for developers just got a cheaper way to sit idle. On November 19, 2026, the OpenAI API changelog added gpt-live-1.1 to the Live API at v1/live/sessions. The entry says GPT-Live-1.1 improves pronunciation, accent consistency and transcription in multilingual voice conversations. It also says developers can configure standby for GPT-Live-1 and GPT-Live-1.1 WebRTC sessions to pause inference during inactivity and resume when someone speaks.
The model is not new to users. It has powered voice in ChatGPT since mid-October, as covered in our post on GPT-Live-1.1, Quip and usage limits by plan. What is new is that API developers can now build on it, and that the billing has a second, lower rate.
TL;DR: what developers need to know
| Question | Answer (OpenAI docs) |
|---|---|
| New model | gpt-live-1.1 on v1/live/sessions |
| What improved | Pronunciation, accent consistency, transcription in multilingual conversations |
| New control | Standby for WebRTC sessions, for GPT-Live-1 and GPT-Live-1.1 |
| Active rate | $0.05 per minute, billed per second |
| Standby rate | $0.02 per minute, billed per second |
| Startup charge | 15 seconds of voice duration on WebRTC creation, credited once the session runs |
| Backend costs | Billed separately by model and tool |
| Not published | New benchmark scores or accent error rates |
What GPT-Live is, in short
GPT-Live separates the voice conversation from the backend that reasons and runs tools. The voice layer talks with the user. When a task needs real work, it delegates to a backend model, which is why the cost guide says to estimate "these two costs separately": the voice session depends on duration, while backend costs depend on the models and tools you use. The changelog says to describe, in plain language, when to delegate to your backend.
That design is different from the older Realtime API, which is a speech-to-speech model priced per token. The pricing page lists gpt-realtime-2.1 with audio input at $32 and output at $64 per million tokens, while GPT-Live is a per-minute rate. They are separate products, so moving between them is a migration, not a flag.
We covered the first API launch in GPT-Live-1 in the API for voice agents. Developers have since built on it; see HeyGen's LiveAvatar open-source demos with GPT-Live-1.
What is better in 1.1
OpenAI describes three improvements: pronunciation, accent consistency and transcription, all in multilingual conversations. A recent OpenAI developer post titled "Prompting GPT-Live for languages and regional accents" appears in the docs showcase list, which suggests accent control is a focus. The changelog does not give numbers. There is no word error rate, accent-similarity score or listener study in the entry we read, so these are vendor statements without published measurement. If your product depends on a specific language or accent, test it with your own recordings.
Standby: how it works and what it saves
Standby is the practical change. In a WebRTC session configured for standby, inference pauses during inactivity and resumes when someone speaks. The cost guide explains the billing:
- Voice sessions accrue charges per second at the active and standby rates.
- Active time includes speech, silence before the idle timeout expires, and time spent waiting for backend work while inference is running.
- The standby rate applies while inference is paused.
- Muting the microphone leaves the session active, so it does not save money.
The example in the docs: a call with 60 seconds active and 30 seconds on standby costs one minute at the active rate plus half a minute at the standby rate. At the listed prices that is $0.05 plus $0.01, or $0.06 of voice cost, before any backend charges.
| Scenario | Active time | Standby time | Voice cost (listed rates) |
|---|---|---|---|
| Always active, 10 minutes | 10 min | 0 | $0.50 |
| Half the time paused | 5 min | 5 min | $0.35 |
| Mostly paused (80%) | 2 min | 8 min | $0.26 |
These are our arithmetic from the published rates, not figures from OpenAI. Standby saves 60% per minute against active time, but only for minutes where nobody is speaking and the session is configured to pause.
Standby fits a kiosk, in-car assistant or ambient agent that stays connected for a long time but is spoken to rarely. For work that runs a long time in the backend, the docs suggest a different pattern: close the voice session while the backend runs a long task, such as coding in goal mode, and start a new session when the result is ready. They show a session.start event that seeds a new session with saved context and the verified task result. Closing stops voice charges, but you pay in reconnection time and a possible interruption.
Cost traps to avoid
The cost guide lists details that change real bills:
- Startup charge. With default startup behavior, creating a WebRTC session with
POST /v1/live/sessionsbills 15 seconds while it initializes. This is credited against duration charges once the session starts, so a 90-second active session already includes those 15 seconds. It matters for apps that reconnect often or create sessions before the user is ready to speak. - Usage snapshots. GPT-Live reports cumulative voice usage in seconds through
session.usage.updated. Each update replaces the previous total. Do not sum the snapshots. Aftersession.close, wait forsession.closedand record its finalusage.secondsonce. - Backend cost. Include model input and output tokens, cached input where supported, and tool charges. A cheaper backend model can cost more overall if it takes longer, repeats tool calls or fails the task, because voice minutes keep running while it waits.
- Speculative work. Starting a lookup from transcript fragments can cut waiting time, but count the unused work in your backend costs.
The cost guide's rule of thumb is to compare cost per successful task, not per minute, and to include retries.
How to try it
- In an existing Live session configuration, set the model to
gpt-live-1.1. The changelog says you can use it in an existing configuration, so migration fromgpt-live-1is a model-name change in principle; test before you ship. - For WebRTC apps, configure standby as described in the Live conversations guide.
- Add usage logging from
session.usage.updatedand the finalsession.closedevent. - Run your own evaluation. OpenAI points to a voice agent evaluation cookbook for comparing task success and cost.
- Compare against other voice stacks before you commit; see our look at Rime Prism's 300 ms claim against its published 570 ms and Cartesia's managed agents pricing claims.
GPT-Live or Realtime: which one for which job
OpenAI's pricing page now lists two voice families side by side, and they suit different products.
| Question | GPT-Live (v1/live/sessions) | Realtime API (gpt-realtime-2.1) |
|---|---|---|
| Billing | Per second of voice session, active or standby | Per token, accrued when a response is created |
| Backend work | Delegated to a separate model you choose and pay for | Same model handles speech and tools |
| Idle cost | Standby rate for paused WebRTC sessions | No connection charge; tokens only when responses run |
| Cost shape | Predictable per minute, plus backend | Grows with conversation length, since the full conversation is re-sent each turn |
Neither is always cheaper. A short, tool-heavy call may cost less on a per-minute plan, while a long session with long silences may cost less per token. Build a spreadsheet from a few real calls before you decide, and include cached-input discounts on the Realtime side.
When standby is the wrong tool
Standby pauses inference, not the session. It helps when users go quiet but stay connected. It does not help when:
- The microphone is muted. The docs state that muting leaves the session active.
- The backend is working. Time spent waiting for backend work while inference is running counts as active time. If your agent often runs long tool calls, close the voice session and restart it, as the docs suggest for long tasks.
- Quick turn-taking matters more than cost. A paused session has to resume when someone speaks. Test the delay on your network before you put it in front of customers.
- You reconnect often. Each new WebRTC session bills the 15-second startup amount, which is credited back only if the session runs.
What is not stated
- Quality numbers. No benchmark, word error rate or accent score is given for 1.1 against 1.
- Standby timing. The idle timeout and the resume latency are in the docs; the changelog entry does not say how long the first word after standby takes. Measure it before you rely on it for fast turn-taking.
- Non-WebRTC sessions. The changelog entry names WebRTC sessions only for standby. WebSocket and SIP behavior is not covered by that line.
- Regional or plan limits. Nothing in the entry restricts availability, but check your rate limits.
What to do now
Builders: Switch a staging session to gpt-live-1.1, turn on standby, and measure resume latency and billing from the usage events.
Founders: Rebuild your cost model on active minutes, standby minutes and backend usage, not on one per-minute number. Idle-heavy products gain most.
Marketers and support leads: Test accents and names your customers actually use. OpenAI's own claim is "more consistent," not a measured gain.
