[KongchangAI]
· 2 min read· 1,177 words

OpenAI Releases GPT-Live-1.1: Voice Agent Idle Cost Drops to $0.02/Minute

OpenAI Releases GPT-Live-1.1: Voice Agent Idle Cost Drops to $0.02/Minute

OpenAI launches GPT-Live-1.1 full-duplex voice model, improving accents and transcription while cutting idle billing to $0.02/min.

OpenAI released GPT-Live-1.1 in the API, a full-duplex voice model that improves multilingual accent handling and transcription quality and adds standby mode, lowering WebRTC idle billing from $0.05 to $0.02 per minute. For voice agent developers, the cost-structure change may matter more than the capability gains: idle unit costs more than halve, making long-lived keep-alive affordable and enabling more “always-on” products; accent improvements expand use in global, cross-border customer service and multilingual education. Active-period billing is undisclosed, so savings depend on session characteristics.

OpenAI has launched GPT-Live-1.1 in the API, a full-duplex voice model. Compared with its predecessor, it improves multilingual accent handling and transcription quality, and adds a standby mode that lowers idle billing for WebRTC connections from $0.05 per minute to $0.02 per minute. For developers building voice agents, the cost-structure change brought by this update may be more noteworthy than the improvement in model capabilities themselves.

OpenAI Ships GPT-Live-1.1 for Voice Agents, Cuts Idle Cost to $0.02/Min

Capability Enhancements for a Full-Duplex Voice Model

The core positioning of GPT-Live-1.1 remains full-duplex voice interaction. Unlike the traditional serial pipeline of “record—recognize—generate—synthesize,” a full-duplex model allows users to keep speaking while the model outputs speech, making the interaction closer to the rhythm of a real human conversation. Interruptions, interjections, and tonal responses can all be handled within the same model.

This update focuses on two areas: multilingual accent handling and transcription. Accent issues have always been a hidden obstacle when deploying voice agents—models perform well on standard pronunciation, but recognition accuracy and conversational fluency drop noticeably with non-native accents and regional pronunciation variants. GPT-Live-1.1 has been optimized for this, and for voice products that need to cover multilingual, multi-region users, this is a more practical capability improvement than simply scaling parameters. Improvements in transcription quality also affect downstream steps: voice agent logging, session analysis, and compliance auditing all depend on transcribed text, and higher transcription accuracy directly reduces the cost of manual review.

The difference between full-duplex and traditional voice pipelines

Traditional voice interaction uses a cascaded pipeline: VAD (voice activity detection) determines whether the user has finished speaking, ASR transcribes to text, the LLM generates a reply, and TTS synthesizes speech. The problem with this approach is that each stage adds latency, and the model cannot “listen” while the user is speaking, so interruption handling requires extra logic. A full-duplex model integrates listening and speaking into the same model, allowing it to continuously receive input while outputting audio, thereby enabling natural interjections and immediate responses. The tradeoff is that the model must model two audio streams simultaneously, which places higher demands on inference infrastructure and latency optimization.

The role of WebRTC in voice agents

WebRTC is the underlying protocol stack for real-time audio and video communication in browsers and mobile apps, providing low-latency audio transmission, echo cancellation, and network jitter buffering. Voice agents typically establish long-lived connections through WebRTC, sending the user’s audio stream to the model and receiving synthesized speech. While the connection is maintained, even if no audio is being transmitted, the server still needs to maintain session state and reserve resources—this is precisely the technical basis for idle billing.

Standby Mode: Pricing “Waiting” Separately

The cost structure of voice agents is fundamentally different from that of text APIs. In a voice session, the user may remain silent for a long time—thinking, looking up information, or walking away—while the WebRTC connection remains open and the model remains on standby. Under traditional time-based billing, this period in which “nothing happens” also consumes charges.

The standby mode introduced by GPT-Live-1.1 prices this idle waiting separately: $0.02 per minute, compared with the previous billing of $0.05 per minute. In other words, the unit cost of idle periods has fallen by more than half. For scenarios such as customer service, voice assistants, and real-time translation, where session durations fluctuate widely, idle time often accounts for a considerable share, and this adjustment will directly change the unit economics of products.

It should be noted that the source material does not disclose the billing method for active interaction periods, so the actual savings from standby mode depend on the ratio of idle time to active time in a given session. Developers should re-estimate costs based on their own business’s session characteristics before launch.

Unit economics

Unit economics refers to the revenue and cost structure of a single user or a single session. For voice agents, key variables include average session duration, the proportion of active interaction, per-minute model invocation cost, and the number of concurrent connections. Lower idle billing directly reduces the variable cost per session, but the savings are significant only when idle time accounts for a relatively high share of total time. If sessions consist mainly of continuous conversation, billing for active periods is the dominant cost driver.

Practical Impact on Voice Agent Developers

The significance of the cost reduction lies not in its absolute value, but in the fact that it lowers the psychological threshold for “keeping the connection open.” Previously, many teams would proactively disconnect after detecting silence and rebuild the session after the user triggered it again in order to control costs, at the expense of interaction continuity and response speed. Standby mode makes long-lived keep-alive affordable, so product design can lean more toward an “always-on” experience.

Improvements in multilingual accents also expand the population that voice agents can serve. For scenarios such as products going global, cross-border customer service, and multilingual education tools, unstable accent recognition in the past often forced teams to narrow their service scope or add manual fallbacks. Model-level improvements reduce the need for such compromises.

From a competitive perspective, full-duplex voice models are becoming a standard capability among API providers, and the focus of differentiation is shifting from “can it converse” to “how low is the conversation cost and how broad is the accent coverage.” The direction of the GPT-Live-1.1 update reflects exactly this shift.

Things to Confirm Before Launch

Before integrating, developers are advised to verify three things: first, whether transcription accuracy for the target languages and accents meets business requirements, preferably tested with real user recordings rather than standard corpora; second, the trigger conditions and billing boundaries of standby mode, clarifying which states count as idle; third, compatibility with existing WebRTC pipelines, to avoid having to rewrite session management logic due to billing model changes.

GPT-Live-1.1 is now available in the API. For teams evaluating voice agent solutions, the value of this update is that it lowers both the capability threshold and the cost threshold at the same time, making more scenarios commercially viable.

One-line Summary

OpenAI launches the GPT-Live-1.1 full-duplex voice model, improving accents and transcription and cutting idle billing to $0.02 per minute.

Paragraph Summary

OpenAI has introduced the GPT-Live-1.1 full-duplex voice model in the API, focusing on improving multilingual accent handling and transcription quality and adding a standby mode that lowers idle billing for WebRTC connections from $0.05 to $0.02 per minute. For voice agent developers, the cost-structure change may be more noteworthy than the capability improvements: unit costs during idle periods have more than halved, making long-lived keep-alive affordable and allowing products to lean toward an “always-on” experience; accent improvements also expand applicability in scenarios such as going global, cross-border customer service, and multilingual education. The source material does not disclose billing for active interaction periods, so actual savings need to be estimated based on business session characteristics. Before integrating, developers should verify transcription accuracy for target accents, standby trigger conditions and billing boundaries, and compatibility with existing WebRTC pipelines.

#OpenAI#GPT-Live-1.1#voice agents#full-duplex#WebRTC#idle billing
Share:

Related articles