Answers from the documentation; checking which model is on duty
Hi - I answer from the OpenSmartRoute documentation: routing, the API, plans and quotas, self-hosting. Ask away, or open a support ticket if you need a person.
Grounded in the docs - follow a source before acting on it.
OpenAI releases GPT-Live-1.1 voice model with standby mode cuts idle c - OpenSmartRoute
OpenAI releases GPT-Live-1.1 voice model with standby mode cuts idle cost
OpenAI has launched GPT-Live-1.1, a full-duplex voice model with a cheaper standby mode, reducing idle session costs. It supports multilingual conversations and native speech processing.
Key points
GPT-Live-1.1 is a full-duplex voice model for multilingual conversations.
Standby mode reduces idle WebRTC billing from $0.05 to $0.02 per minute.
Model supports a 128,000-token context window during live sessions.
Released in API as gpt-live-1.1 on November 21, 2026.
Why it matters: Cost savings and improved efficiency for voice agents through lower idle session billing and native speech processing.
By OpenSmartRoute editorial · written through the router by llm-small
From MarkTechPost - “OpenAI Ships GPT-Live-1.1 for Voice Agents, Cuts Idle Cost to $0.02/Min”
Introduction - announcement of GPT-Live-1.1 and its main features
OpenAI has released GPT-Live-1.1, a new voice model for its API. This model is designed for full-duplex conversations, meaning it can listen and speak at the same time. It introduces a standby mode that lowers idle session costs. The release aims to improve voice interactions by making them more natural and cost-efficient. It is now available under the model ID gpt-live-1.1.
GPT-Live-1.1 supports multiple languages and native speech processing. It is built to handle real-time voice conversations better than previous models. The model is used mainly for voice agents, chatbots, and other speech-based applications. It runs only on OpenAI’s cloud infrastructure and connects via WebRTC or WebSocket. This update marks a step forward in making voice AI more practical and affordable for businesses.
Model capabilities - full-duplex speech, multilingual support, and context window
GPT-Live-1.1 is a full-duplex model, which means it can listen and speak simultaneously. This allows for more natural conversations, as the model can respond while still hearing the user. It processes audio and text inputs and produces audio and text outputs. Unlike traditional speech systems, it reasons over incoming and outgoing audio in a single process.
The model supports multiple languages, making it suitable for global applications. It can handle conversations in different languages without switching models. The context window for live sessions is 128,000 tokens by default. This large window helps the model remember and process long conversations, improving its ability to maintain context over time.
Performance and improvements - accents, transcription, and task handling
OpenAI claims that GPT-Live-1.1 offers better accents and speech quality across languages. It also improves transcription accuracy, especially for letters and numbers. The model is designed to handle detailed instructions more reliably. It aims to complete tasks with greater consistency and fewer errors.
While no official benchmark scores are published, OpenAI states that GPT-Live-1.1 performs better than previous versions in these areas. The model can also delegate reasoning and tool calls to backend text models. This allows it to handle complex tasks or high-volume workloads more efficiently. It supports streaming and function calling, but not image or video outputs.
Countries agreed to start discussing options for rules around autonomous weapons after a week of talks. No binding rules were set.
Standby mode - how it reduces idle cost and configuration options
The standby mode is a new feature that reduces costs when a voice session is idle. When enabled, the model pauses inference during silence, but the WebRTC connection remains open. This means the model stops actively processing audio but stays ready to resume quickly when the user speaks again.
To activate standby mode, users set specific parameters. The wake policy type is set to speech, so the model waits for speech before responding. The idle timeout can be configured from 5 to 60 seconds. After this period of silence, the server emits a session.paused event. When the user speaks again, a session.resumed event fires, and the conversation continues seamlessly.
This feature helps lower billing costs. It is available for both GPT-Live-1 and GPT-Live-1.1 models. The server keeps the session ID and history during pauses, ensuring smooth transitions. However, the first reply after a pause may be slightly slower, and users need to keep the microphone active during standby.
Pricing and billing - session costs, separate backend charges, and API details
Voice sessions are billed based on their duration. For GPT-Live-1.1, billing is calculated per second, making it precise for short or long calls. When using standby mode, idle WebRTC time costs less—at a rate of $0.02 per minute—compared to active sessions at $0.05 per minute. This results in a 60% reduction in idle session costs.
Backend processing, such as reasoning or tool calls, is billed separately from voice minutes. This means that a heavy backend workload can increase overall costs beyond the voice session charges. The API is the only way to access GPT-Live-1.1; switching models requires starting a new session. There are no open weights or quantized variants available for this model.
Pricing details are published on the model page, and users can check deployment options and specifications through tools like the GPULeaderboard.com chrome extension. The billing structure emphasizes cost efficiency, especially when using standby mode for long or idle conversations.
Deployment options - connection methods, session management, and limitations
Clients connect to GPT-Live-1.1 via WebRTC or WebSocket, both of which support real-time audio streaming. The model runs exclusively on OpenAI’s cloud infrastructure, and there are no open-source weights or quantized versions available. This means users cannot run the model locally or modify it directly.
Managing sessions involves setting the model ID to gpt-live-1.1 and handling session events such as paused and resumed. Each session is independent, and changing the model requires starting a new session. The model supports outbound SIP calling, which was made available to all organizations on November 2, enabling phone-based voice agents.
There are limits on concurrent sessions, but OpenAI does not publish specific tier numbers for GPT-Live-1.1. The model is designed for real-time, continuous conversations, with configuration options for wake policies, timeouts, and silence detection. Users should consider these limits when planning large-scale deployments.
Background - what GPT-Live models replace and how they work
GPT-Live models replace the traditional speech-to-text, language model, and text-to-speech pipeline. Instead of separate components, GPT-Live reasons over audio and text together in a single model. This simplifies the architecture and reduces latency.
The models handle turn detection and return automatic speech recognition (ASR) transcripts natively. They can delegate complex reasoning or tool calls to backend models, such as GPT-6 Astra for complicated issues or Luna for high-volume tasks. Third-party backends can also be integrated.
GPT-Live models are designed for live conversations, guiding the flow based on user input and system responses. They support streaming outputs and function calls, but do not support images, videos, or structured outputs. The models are API-only, and their design aims to make voice interactions more seamless and integrated.
Why it matters - benefits for voice agents and cost efficiency
The release of GPT-Live-1.1 brings several benefits for voice agents and conversational AI. Its full-duplex capability allows for more natural, human-like interactions. Support for multiple languages broadens its usability across different regions.
The standby mode significantly reduces idle session costs, making long conversations more affordable. This is especially useful for applications with frequent pauses or long periods of silence. The ability to delegate complex reasoning to backend models improves reliability and task handling.
Cost efficiency is a key factor. Lower idle billing rates and the ability to keep sessions open without active inference save money. This makes GPT-Live-1.1 suitable for large-scale deployments where managing expenses is critical. It also opens new possibilities for real-time voice applications that require continuous listening and speaking.
How OpenSmartRoute helps - routing voice sessions through the platform
OpenSmartRoute is an open-source router designed for AI requests. It manages how requests are sent to models, agents, or tools from a defined catalog. It scores each candidate based on quality, cost, speed, and safety, with team-set weights.
Routing through OpenSmartRoute offers cost savings and better control. It learns from outcomes, directing more traffic to well-performing models. When a new model like GPT-Live-1.1 is introduced, it can be added to the catalog and tested against existing options. The platform ensures requests stay within rules, such as data privacy and regional limits.
Using OpenSmartRoute with GPT-Live-1.1 allows teams to optimize voice session costs. They can route requests to the most suitable model based on the conversation context and desired quality. The platform also provides visibility into costs and performance, helping teams manage large-scale voice deployments efficiently.
How it compares - what existed before, what this changes and what stays the same
Before GPT-Live-1.1, OpenAI offered GPT-Live-1, which was a speech model designed for real-time conversations. It could listen and speak, but it did not support standby mode. The previous model used separate speech-to-text, language understanding, and text-to-speech systems. This setup required multiple steps and added latency.
GPT-Live-1.1 introduces a full-duplex model that reasons over audio streams in a single process. It handles turn detection, transcriptions, and speech synthesis all at once. This change simplifies the architecture and reduces the delay between listening and speaking. The model also improves accents, transcription accuracy, and task completion, according to OpenAI.
The main features that stay the same include streaming outputs, function calling, and API-only access. The model still does not support images, videos, or structured outputs. It continues to be designed for live voice conversations and can delegate complex reasoning to backend models. The core purpose remains: making voice interactions more natural and integrated.
What changes significantly is the addition of standby mode. This feature allows sessions to pause inference during silence, reducing costs. It keeps the session active, but no inference runs until the user speaks again. The default context window remains large at 128,000 tokens, supporting long conversations. The billing model now includes a cheaper idle rate, making continuous listening more affordable.
OpenAI also states that GPT-Live-1.1 improves accents and handling of details like letters and numbers. These are self-reported gains, with no published test scores. The model runs only on OpenAI’s cloud, with no open weights or quantized versions available. It connects over WebRTC or WebSocket, supporting real-time voice applications.
In summary, GPT-Live-1.1 builds on the previous model by adding standby mode, improving naturalness, and simplifying the architecture. It maintains core features like streaming and delegation, while offering cost savings for idle sessions.
Questions this leaves open - what the source does not say and how a reader can check it
The source does not provide benchmark scores or detailed evaluation results for GPT-Live-1.1. It claims improvements in accents, transcription, and task completion but offers no numbers or tests. This makes it hard to compare its performance objectively against other models or earlier versions.
It is unclear how much the standby mode affects latency after user input. The source mentions that the first reply after a short utterance may be slower, but it does not specify how much. Users may want to test this in their own applications to see if the delay is acceptable.
The maximum number of concurrent sessions per region is not published. Teams need to know their limits to plan large deployments. They can check this by contacting OpenAI support or testing in their environment.
The source does not specify the exact pricing tiers or how billing scales with usage. While it states that standby reduces costs, teams should verify the actual rates and billing policies through OpenAI’s documentation or billing portal.
It also does not mention the model’s robustness across different languages or accents. Although it claims more natural accents, users should evaluate the model in their target languages and dialects. They can do this by running test conversations and comparing results.
The API documentation details are limited. Teams should review the full API specs, especially for features like wake policies and timeout settings. This helps in integrating standby mode effectively.
Finally, the source does not specify how third-party backends are integrated or how they interact with GPT-Live-1.1. Teams interested in custom reasoning or tool calls should explore OpenAI’s backend delegation options and test their workflows.
In conclusion, while GPT-Live-1.1 offers new features and claimed improvements, users need to conduct their own tests. They should verify performance, latency, costs, and robustness in their specific use cases. Contacting OpenAI or reviewing detailed API docs can help clarify these points.