GPT-Live-1.1 API Release: Voice Updates and Standby Pricing

On this page
GPT-Live-1.1 is now available through OpenAI’s API, bringing updated multilingual voice handling and a standby option that costs $0.02 per minute instead of the $0.05 active rate. My first question is less about how charming an assistant sounds and more about what happens when a caller corrects an address halfway through a sentence. That is where this release deserves attention: conversational accuracy, interruptions, and the handoff between speaking and doing.

OpenAI’s API changelog dates the update to November 19, while its developer-community announcement appeared November 20. This is a newly available API model being examined in our November 21 AI briefing, rather than a claim that OpenAI announced it today. The company says GPT-Live-1.1 had already powered voice in ChatGPT Plus and Pro since October 15. API availability gives application developers a separate route to build with it.
GPT-Live-1.1: the confirmed changes
| Item | Verified detail |
|---|---|
| API model ID | gpt-live-1.1 |
| Conversation behavior | Full duplex: can listen while speaking |
| Active session price | $0.05 per minute |
| Standby price | $0.02 per minute for configured WebRTC standby |
| Idle timeout | Configurable between 5 and 60 seconds |
| Other costs | Backend model and tool usage billed separately |
OpenAI describes improvements in pronunciation, accent consistency, and transcription across languages. Its announcement also highlights understanding precise details such as letters and numbers. Those are vendor claims, not results from a Kocitech listening test. We have not independently measured error rates or compared recorded calls between the two model generations.
The official developer announcement is useful because it separates API access from the earlier ChatGPT rollout. Developers can select the new model ID in an existing Live configuration, but a model change still deserves a regression check. A voice assistant that sounds more fluent can remain wrong about a surname, an order number, or an instruction.
Why the conversation and the backend are separate
GPT-Live handles the spoken exchange while a backend model or service performs delegated work. OpenAI’s getting-started guide describes two paths: Responses delegation, where OpenAI manages the connection to a supported backend model, and client delegation, where the application connects its own agent or service. The voice layer can keep the conversation moving while that work continues.
Think about a delivery query. A caller asks when a parcel will arrive, then adds that the delivery address changed. The system needs to hear the correction, obtain the right order, and avoid confidently speaking an answer based on stale information. Fluent speech is only one part of that chain. The application still needs a clear definition of which information is authoritative and which actions require confirmation.
This division also changes how I would assess an upgrade. A failed database lookup should not automatically be counted as a speech-recognition failure. A correctly heard request followed by an unauthorized account change is a workflow failure. Testing those stages separately makes it easier to identify whether the voice model, backend instructions, integration, or business rules need attention.
Standby savings depend on actual quiet time
Standby pauses inference during inactivity while keeping a WebRTC call open. Speech resumes the conversation. OpenAI makes this available to both GPT-Live-1 and GPT-Live-1.1, so teams should not confuse a billing feature shared by the two versions with an exclusive capability of the new model.
Here is a simple illustration using the published duration rates. Ten active minutes cost $0.50 for the voice session. If a ten-minute session instead contains six active minutes and four minutes billed in standby, the voice portion would be $0.38: $0.30 plus $0.08. That is a $0.12 difference, or 24 percent in this example. It is arithmetic, not a forecast of anyone’s production bill.
The result depends on how much time actually reaches standby after the configured timeout. Brief pauses may never qualify. Backend reasoning, tools, and other service charges remain additional. A sensible cost review should therefore compare complete sessions with the same workload, then examine the voice component separately. Looking only at the cheaper standby figure would understate what an application costs to operate.
What I would test before switching a real service
Start with a small set of calls that reflect the people using the product. Include names spoken with different accents, mixed-language sentences, spelled identifiers, and numbers that sound similar. Keep the expected outcome explicit: which letters should be stored, what information should be repeated, and when the assistant should ask the speaker to clarify.
Next, interrupt it. Correct a street name while it is speaking. Add an item after it starts a backend lookup. Ask it to stop, then resume with a different request. A useful evaluation checks the final saved state as well as the transcript. Otherwise, a polite spoken acknowledgment can conceal a backend task that continued with the old instructions.
Finally, exercise quiet moments. Let a caller look for an order number, then return after the timeout. Check that the application records the session state and resumes correctly. These are proposed evaluation scenarios, not tests Kocitech has run. Our recent Gemini 4 Neon coverage takes a similar approach to separating release claims from workflow validation.
A better voice does not settle permission questions
OpenAI’s prompting guidance recommends keeping conversation instructions short and placing detailed procedures in the backend. That is a useful design boundary. Speaking style belongs close to the conversation; refund rules, account permissions, and required confirmations belong in the application workflow. A convincing voice should not become the authority for deciding what an account holder is allowed to change.
I would pay particular attention to announcements of completed work. The assistant should describe a booking as confirmed only after the booking system reports success. If a tool is slow or fails, the spoken update needs to reflect that uncertainty. Our explanation of OpenAI’s embedded assessment pilots explores the wider distinction between a system’s reported behavior and independently checked evidence.
For developers, the immediate opportunity is a more capable conversational layer with an explicit way to reduce inference during silence. The practical decision remains grounded in actual callers, full session costs, and completed tasks. I would start with a limited evaluation, compare the same scenarios against the existing configuration, and expand only after the results justify it.
Frequently asked questions
Is GPT-Live-1.1 available in the API now?
Yes. OpenAI’s API changelog and developer announcement confirm availability using the model ID gpt-live-1.1. This article covers the API release; it does not imply that ChatGPT first received the model on November 21.
Does standby make every minute cost $0.02?
No. The lower rate applies to configured WebRTC standby periods. Active voice sessions cost $0.05 per minute, and backend model and tool usage is charged separately. Your mix of active and standby time determines the voice bill.
Can it listen while it is speaking?
Yes. OpenAI describes GPT-Live-1.1 as full duplex. Developers should still test how their own application handles interruptions, corrections, and backend work already in progress before relying on that behavior with real customers.
Has Kocitech independently tested the claimed improvements?
No. This is a source-based release analysis. Pronunciation, accent, transcription, and task-reliability improvements are attributed to OpenAI, and the suggested call scenarios are a proposed evaluation plan rather than reported testing results.
