SIAgentNews.com
← All news
DevelopingOpenAIAgents

OpenAI releases GPT-Live-1.1 to lower voice agent costs

The new model introduces a standby mode that cuts idle costs for voice agents by 60% while improving multilingual accents.

· First reported by MarkTechPost

$0.02cost per minute for idle sessions

The short version

  • GPT-Live-1.1 is a full-duplex model that can listen and speak simultaneously in a single stream.
  • A new standby mode reduces idle billing from $0.05 to $0.02 per minute, a 60% cost reduction.
  • The model offers improved accents, transcription accuracy, and handling of specific details like numbers.
  • The model is available via API and supports WebRTC and WebSocket connections for voice agents.

Why it matters

For businesses building automated phone agents or voice assistants, this update makes 'always-on' connections much more affordable. It allows for more natural, human-like interactions without the previous high cost of maintaining a constant connection.

A more natural voice experience

While the model is an improvement over the previous version, OpenAI has not published any specific benchmark scores to quantify these gains. The improvements are currently self-reported by the company. Users can access the model by switching to the gpt-live-1.1 model ID in their existing API setups.

How the model processes audio

Developers can choose between different delegation types depending on their needs. In responses mode, the system uses OpenAI models for reasoning. In client mode, the voice model can connect to a third-party backend. This flexibility allows for customized workflows while maintaining a consistent voice interface.

Cutting the cost of silence

To use this feature, developers must set the wake policy to speech. This tells the server to wait for the user to speak before resuming inference. While the session remains paused, the session ID and history are preserved. OpenAI notes that the first reply after a short utterance might be slightly slower because the model is resuming from a paused state.

Deployment and technical limits

Organizations using outbound SIP calling for phone agents can now utilize this model. OpenAI has established separate concurrent-session limits for GPT-Live and GPT-Live-1.1. These limits apply across different inference regions, though the specific tier numbers are not currently published on the model page.

Historical context and deployment

The release of GPT-Live-1.1 follows the initial launch of the GPT-Live model, which first reached the API on September 10, 2026. This progression shows a steady evolution in how OpenAI handles real-time audio interactions for developers and end users alike.

According to the OpenAI forum announcement, this technology has already been powering the voice features for ChatGPT Plus and Pro users since October 15. This means the underlying architecture has been refined through significant real-world usage before being made available for broader API integration.

Specialized backend models

While the voice model handles the primary interaction, OpenAI uses specific models to manage different types of reasoning tasks. For example, the company names GPT-6 Astra as the designated model for handling complex issues that require deeper logic or multi-step thinking.

In contrast, a model called Luna is used for high-volume tasks where efficiency and scale are the primary requirements. This tiered approach allows the voice layer to delegate work to the most appropriate backend, whether it is an internal OpenAI model or a third-party system.

Technical background and community

The information regarding these updates is supported by data science professionals who specialize in transforming complex datasets into actionable insights. These experts use a solid foundation in statistical analysis, machine learning, and data engineering to understand how these models perform in production environments.

Native audio processing features

The GPT-Live-1.1 model is designed to handle the complexities of human conversation by managing several audio tasks natively. This includes the ability to perform turn detection, which allows the system to understand when a user has finished speaking and when it is time to respond. Additionally, the model returns automatic speech recognition transcripts directly as part of the native output stream.

While the model is highly capable in real-time interaction, it does have specific functional limitations that developers need to consider. The system does not currently support structured outputs, fine-tuning, or predicted outputs. These restrictions mean that while the voice interaction is fluid, certain types of rigid data formatting or custom model training are not available through this specific model ID.

Configurable standby settings

To manage the new standby mode effectively, the documentation provides three specific settings that developers can adjust. The wake policy type can be set to speech, ensuring that the system waits for actual vocal input before it begins inference. Users can also define an idle timeout, which is a whole number that can be set anywhere from 5 to 60 seconds to determine how long the system waits before pausing.

Another optional setting is the voice activity detection threshold, which can be adjusted between 0 and 1 with a default of 0.5. These configurations allow for precise control over how the model behaves during moments of silence. For those looking to collaborate or share their work, there are also opportunities to partner for promoting GitHub repositories, Hugging Face pages, product releases, or webinars.

What to watch

  • How the self-reported accuracy gains for numbers and letters hold up in real-world phone agent deployments.
  • Whether OpenAI will release official benchmark scores for the 1.1 version in the future.

Try it

If you use OpenAI's API for voice, check your current idle costs and see if switching to the gpt-live-1.1 model ID with standby mode can reduce your per-minute billing.

Sources and links

Read the original at MarkTechPost

First reported by MarkTechPost. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.

Liked it? Say why

Add a line for other readers.

No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.

More AI news