OpenAI has made GPT-Live-1.1 available to developers through its API, giving businesses and software developers access to a voice model designed for natural, multilingual, real-time conversations. Announced on November 19, 2026, the release brings improvements in speech transcription, pronunciation and accent consistency, alongside a new standby feature that can reduce the cost of voice applications during periods of inactivity.
The most notable change is the introduction of a lower-priced standby rate for eligible WebRTC voice sessions. Developers can now pay $0.02 per minute while inference is paused, compared with the standard active-session rate of $0.05 per minute. The feature is designed for applications such as customer support agents, virtual assistants and voice-enabled services that need to maintain an open connection even when users are temporarily silent.
Key Takeaways
- OpenAI has released GPT-Live-1.1 through its API for developers.
- The model improves transcription accuracy and maintains more consistent accents and voices across languages, according to OpenAI.
- Active voice sessions cost $0.05 per minute, while eligible standby sessions cost $0.02 per minute.
- Standby mode is available for WebRTC sessions using both GPT-Live-1 and GPT-Live-1.1.
- Developers can configure an inactivity timeout between five and 60 seconds.
- Backend reasoning models and tool usage are billed separately.
- The model supports streaming and function calling but does not support image or video inputs.
GPT-Live-1.1 Is Now Available Through the OpenAI API
GPT-Live-1.1 is a full-duplex voice model, meaning it can listen while speaking. This capability allows developers to build applications that support more natural conversations without requiring users to interact through separate speech-recognition and text-to-speech stages.
The model was already powering voice experiences in ChatGPT Plus and Pro from October 15, 2026. Its November API release extends access to external developers and businesses that want to integrate the technology into their own products.
OpenAI says GPT-Live-1.1 improves on its predecessor, GPT-Live-1, in several areas:
- More accurate speech transcription.
- More consistent pronunciation and accents across languages.
- Better handling of details such as letters and numbers.
- More reliable instruction-following and task completion.
These improvements could be useful for businesses developing multilingual customer service systems, voice-based booking applications, conversational assistants and other real-time communication products.
OpenAI has not published a quantified benchmark in the model listing showing the exact percentage improvement in transcription accuracy. Developers will therefore need to evaluate the model against their own use cases before deciding whether to migrate existing applications.
New Standby Mode Reduces Voice Application Costs
The most commercially significant part of the release is the new standby configuration.
Voice applications often maintain an active connection even when nobody is speaking. For example, a customer may pause before answering a question, wait while an agent retrieves information or remain on hold during a support call.
Previously, developers had fewer options for reducing the cost of these quiet periods without closing the voice session.
OpenAI’s standby feature allows eligible WebRTC sessions to pause inference after a configured period of inactivity. The connection remains open, and the conversation can resume when the user speaks again.
This separates the cost of maintaining a session during inactivity from the cost of actively processing a conversation.
GPT-Live pricing comparison
| Session type | Price per minute |
|---|---|
| Active voice session | $0.05 |
| Standby voice session | $0.02 |
| Saving during standby | 60% |
The standby rate is 60% lower than the active rate. This calculation applies to the voice-session duration charge, not the total cost of running an AI application.
Backend model usage and tool calls are billed separately.
How much could developers save?
Consider a voice application that maintains a session for one hour.
| Usage scenario | Voice-session cost |
|---|---|
| 60 minutes fully active | $3.00 |
| 60 minutes fully on standby | $1.20 |
| 30 active + 30 standby minutes | $2.10 |
These examples use OpenAI’s listed per-minute rates and exclude backend reasoning, tool calls and other application expenses.
For a business operating thousands of voice sessions, the savings could become meaningful if a substantial share of session time is spent waiting or inactive.
However, standby does not automatically apply to every quiet moment. Developers must configure the feature, and the lower rate applies when inference is actually paused in a supported WebRTC session.
Developers Can Configure Standby for GPT-Live-1 and GPT-Live-1.1
An important detail is that standby mode is not exclusive to the new model.
OpenAI has made the configuration available for both GPT-Live-1 and GPT-Live-1.1 WebRTC sessions. This means developers using the earlier model may also be able to reduce costs without immediately migrating to GPT-Live-1.1.
The inactivity timeout can be configured between five and 60 seconds. Once the selected period of silence has passed, inference pauses while the WebRTC connection remains open.
When speech resumes, the conversation can continue.
This is particularly relevant for applications in which disconnecting and reconnecting would create a poor user experience.
Examples include:
- Customer support: A voice agent can remain connected while a customer checks information or waits for assistance.
- Virtual receptionists: Businesses can maintain conversational sessions without paying the active rate throughout every quiet period.
- Automotive assistants: Voice interfaces can handle natural pauses between commands.
- Healthcare administration: Appointment and information services can support conversations with longer pauses, subject to appropriate privacy and compliance safeguards.
- Enterprise assistants: Employees can interact with voice-based systems without requiring the application to restart after every pause.
The practical savings will depend on how much time applications spend in standby and how developers configure their session behaviour.
Backend AI Costs Are Still Separate
Although the standby rate reduces the cost of eligible voice-session time, it does not make the entire application cheaper by the same percentage.
GPT-Live separates the voice interaction layer from the backend system responsible for more complex reasoning and tool use.
The voice model manages the real-time conversation. When a user asks for something that requires additional processing, the application can delegate the task to a backend model or connected tool.
For example, a customer might ask a voice assistant to check an order. The voice model handles the spoken interaction, while the backend retrieves the order information from the company’s system.
These backend operations carry separate charges.
Developers therefore need to calculate at least two components:
- Voice-session expenses: Based on active and eligible standby duration.
- Backend expenses: Based on the reasoning models, tokens, tools and external services used.
OpenAI’s cost optimisation guidance also explains that developers should account for the time spent waiting for backend work while inference remains active. Simply muting the microphone does not put a session into standby.
This distinction matters when estimating the actual operating cost of a production voice agent.
GPT-Live-1.1 Supports Streaming and Function Calling
The model is designed specifically for real-time voice applications rather than serving as a general-purpose replacement for every OpenAI API model.
Its documented capabilities include:
| Capability | GPT-Live-1.1 |
|---|---|
| Audio input | Supported |
| Audio output | Supported |
| Text input and output | Supported |
| Real-time streaming | Supported |
| Function calling | Supported |
| Image input | Not supported |
| Video input | Not supported |
| Structured outputs | Not supported |
| Fine-tuning | Not supported |
Developers access the model through the Live API endpoint, v1/live/sessions, using the model ID gpt-live-1.1.
Function calling allows voice applications to interact with external systems and execute tasks. This could include retrieving customer records, checking availability, updating bookings or connecting to business software.
However, developers should understand the model’s limitations before choosing it for an application. It does not support image or video inputs, structured outputs or fine-tuning.
OpenAI also lists separate concurrent-session limits for GPT-Live-1 and GPT-Live-1.1. Businesses planning large deployments should check their organisation’s applicable rate limits before launching.
What the Release Means for AI Voice Agents
The release comes as businesses increasingly explore voice agents for customer service, sales, internal support and automated business processes.
Traditional automated phone systems often depend on fixed menus and predefined commands. Modern AI voice agents can understand natural language, respond conversationally and connect to backend systems to complete tasks.
However, operating these systems at scale requires more than a capable voice model. Businesses must also manage latency, reliability, infrastructure costs and the quality of the underlying data.
The new standby pricing addresses one part of that challenge.
A voice agent that remains connected for long periods may spend a considerable amount of time waiting for a user to respond. Reducing the cost of those inactive periods could make certain always-available voice applications more economical.
The benefit may be particularly relevant to companies operating customer support services, appointment systems and conversational interfaces with unpredictable call durations.
The model’s multilingual improvements could also help developers build products for users who speak different languages or expect more consistent pronunciation during conversations.
However, actual performance will depend on the language, accent, background noise, application design and backend systems involved.
Why This Matters for Developers and Businesses in India
India is an important potential market for AI-powered voice applications because businesses frequently interact with customers across multiple languages and regions.
Banks, telecom operators, e-commerce companies, travel platforms and service providers could explore voice agents for customer support, account assistance, order updates and appointment management.
More consistent accents and improved transcription could help developers build applications for multilingual environments. However, OpenAI’s announcement does not establish a quantified improvement for any specific Indian language, so developers will need to test performance in the languages and accents their customers actually use.
The lower standby rate could also help Indian startups experimenting with AI-powered calling products and automated support systems. Lower session costs may improve the economics of applications that remain connected during customer pauses.
Nevertheless, the overall cost will depend on backend model selection, tool usage, call duration and the amount of time spent in standby. Businesses will also need to consider data protection, consent requirements and applicable rules governing automated calls.
For Indian AI startups, the release provides another option for developing real-time voice products without building the entire speech interaction infrastructure independently.
The Bigger Picture
OpenAI’s GPT-Live-1.1 release reflects a broader shift in the AI industry: competition is increasingly moving beyond model intelligence towards the economics of deploying AI systems in real-world products.
For voice applications, developers need models that can understand speech, respond naturally, maintain conversational continuity and connect with external business systems. But they also need predictable operating costs.
The standby feature addresses a specific but important expense: paying the active inference rate during periods when users are not speaking. By introducing a lower rate for eligible paused WebRTC sessions, OpenAI gives developers another way to manage the cost of persistent voice connections.
The release also illustrates how AI providers are separating different parts of the application stack. Voice interaction, backend reasoning and tool execution can have different cost structures, allowing developers to optimise each component independently.
Looking Ahead
Developers will now be able to evaluate GPT-Live-1.1 against their existing voice systems, particularly for multilingual transcription, accent consistency and real-time task execution. Businesses already using GPT-Live-1 may also examine the standby configuration because it is available for both model versions. The key question will be whether the lower session costs translate into meaningful savings without compromising response speed or user experience.
As AI voice agents become more common in customer service and enterprise software, pricing flexibility could become an important factor in choosing an API provider. OpenAI’s latest update does not eliminate the separate expenses associated with backend reasoning and tools, but it gives developers a more economical way to manage inactive voice sessions. The longer-term opportunity will depend on how reliably companies can turn these capabilities into useful, cost-effective products.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



