OpenAI brings GPT-Live-1.1 to developers with lower-priced standby sessions
OpenAI says the voice model improves transcription and multilingual consistency. Paused sessions qualify for a lower rate, but reasoning and tools carry separate charges.
Loading page…
OpenAI says the voice model improves transcription and multilingual consistency. Paused sessions qualify for a lower rate, but reasoning and tools carry separate charges.
Listen to this story
OpenAI has opened GPT-Live-1.1 to developers through its API, turning a model previously documented as the successor to GPT-Live-1 into an option for deployed voice applications. Pricing is $0.05 per active minute and $0.02 per minute only for eligible WebRTC sessions while inference is paused; backend reasoning and tools cost extra. New safety evaluations cover the voice model with delegation disabled, so they do not establish real-world unsafe-response rates. Developers also need to account for separate concurrency limits and unsupported structured outputs, fine-tuning, images, and video.
GPT-Live-1.1 supports streaming and function calling, but not structured outputs, fine-tuning, image inputs, or video.
The $0.02-per-minute standby rate applies only to WebRTC sessions while inference is paused; quiet moments alone do not qualify.
Backend reasoning and tool calls are billed separately, and developers define when the voice model delegates work.
Developers can now put OpenAI’s GPT-Live-1.1 into voice applications that listen while speaking. OpenAI says the API model improves transcription and keeps accents more consistent across languages. Sessions cost five cents per active minute, with a lower standby rate for qualifying pauses—but backend reasoning and tools remain a separate bill.
GPT-Live-1.1 already had an October 15, 2026 system card describing it as the successor to GPT-Live-1. Developer API access is the change now. OpenAI’s November 19 addition to that card also includes evaluations of the API voice models, extending the document beyond its earlier account of the family.
The model retains full-duplex conversation: it can receive speech and generate speech at the same time. OpenAI describes that design as a way to follow pauses, interruptions and changes in pace, deciding whether to answer or continue listening rather than requiring each speaker to take a strictly separate turn.
For this version, OpenAI’s stated improvements focus on what the model hears and how it sounds. The company says it transcribes more accurately than GPT-Live-1 and maintains a steadier accent and voice across languages. The model page does not quantify either improvement or publish a speed figure.
OpenAI bills voice sessions by duration, with separate active and standby prices. The two-cent standby rate applies specifically to WebRTC sessions while inference—the model’s processing—is paused. It is not a blanket price for every quiet moment or every kind of connection; eligibility depends on that paused session state.
The division of work is explicit. Developers describe in plain language when GPT-Live should hand a request to their backend for reasoning or tool use. The voice model handles listening and speaking, while a developer-chosen backend model can take requests that need more reasoning. That backend’s usage sits outside the posted voice-session rate.
The updated system card evaluates both API voice models with backend delegation disabled. OpenAI says this isolates their own safety behavior, rather than letting the backend reasoning model determine the result. In an application that uses delegation, the backend’s safeguards apply to the work it performs.
On tests built from user-contributed audio, OpenAI says GPT-Live-1.1 API scores numerically higher than its predecessor on mental health, emotional reliance and self-harm. Sexual content, illicit behavior and personal data scores are at parity. These are deliberately difficult tests, not estimates of how often unsafe responses occur in everyday use.
The API model also scores numerically higher across all eight synthetic safety categories. OpenAI cautions that those constructed scenarios target different risks and may not predict production behavior. The results describe the voice model under a particular test setup, not a complete delegated application.
For developers moving from GPT-Live-1, capacity is another separate consideration: the two models have independent concurrent-session limits. Each model’s limit applies across inference regions, and OpenAI directs developers to their organization’s limits for the allowance on their account.
Loading discussion...
Join the conversation
Explain when testing the voice model alone is useful—or misleading.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.