On September 10th, OpenAI opened up the service from GPT – Live to API, pricing it at $0.05 per minute for the front-end voice layer. It uses full-duplex technology to process audio, allowing users to listen and speak simultaneously. It also enables the transfer of complex reasoning and tool calls to backend text models such as GPT – Astra as the conversation continues. For developers, this is not just about replacing the voice with a more natural one; it's about transforming the voice system from a serial process of "recognition – reasoning – synthesis" into a structure where front-end dialogue and backend execution work in collaboration.
The most obvious problem with traditional voice robots is not that they can't hear each word clearly, but rather that they don't know when to speak and when to stop. If the user pauses for a moment, the system might answer prematurely; if the user changes their statement midway, the already generated response will continue to play; and if there is someone talking in the background, the robot may mistake their voice for a command. Each transition between voice to text, model response, and text to voice adds latency, and it's also easy to lose the tone, pauses, and interruption signals.
Full-duplex reduces the number of mechanical cycles, but the test results do not guarantee stability in all scenarios.
GPT – Live – 1 uses the same model to jointly process input and output audio, with the official emphasizing its capabilities in interrupt handling, background noise management, silence management, and reliability for long conversations. Speak stated in early evaluations that compared to previous round-based systems, the number of interruptions during language learners' thinking was reduced by nearly 80%. Another client reported that after switching to this model, the amount of code was reduced by 80%, with about 23,000 lines of code being deleted. It performs 30 percentage points better than GPT – Realtime – 2.1 on Full Duplex Bench, and can be combined with backend models to handle end-to-end tasks.
These numbers indicate that the architecture has potential benefits, but they are all subject to the limitations of the testing environment and customer implementation. The patterns of pauses in language learning are different from those in emergency hotlines, bank customer service, or noisy restaurants; the code that a company decides to delete also depends on how many intermediate components are present in the original system stack. Development teams cannot simply write “80% reduction” into their business commitments; they must re-measure these effects using real accents, devices, networks, and business scripts.
Full-duplex also changes the nature of errors. Serial connection systems, although slower, are easier to diagnose whether the error lies with recognition, the model, or synthesis; end-to-end voice models are more natural, but they require simultaneous recording of the audio timeline, transcription, system actions, and the points where interruptions occur. When a customer says “do not cancel,” if the agent has already submitted the cancellation request to the backend, the cessation of sound does not mean that the action has stopped. The front-end dialogue state and the backend transaction state must be managed separately.
The official provides more voice options for accents, dialects, and languages, and also allows for adjustments to tone, rhythm, and style through system prompts. To customize a voice, one still needs to contact sales and meet the qualifying conditions; it should not be assumed that all developers can immediately clone any voice. Enterprises should also clearly inform users that they are communicating with AI, prohibit unauthorized imitation of real people, and set retention periods for recorded audio, voiceprint data, and transcribed content.
For a real launch, it's necessary to take into account latency, permissions, the need for manual intervention, and costs all together.
Voice agents are suitable for appointments, order inquiries, and general customer service, as the issues are usually structured and the backend actions can be clearly confirmed. When designing the process, it is important to separate "understanding the intent" from "executing the action." For inquiries about business hours, a direct response is possible; for modifying an address, a repetition for confirmation is needed. However, actions such as payment, cancellation, or medical arrangements require stronger verification. If a user interrupts once, it should not automatically be considered as consent or cancellation. Critical actions must be completed with clear questions to ensure a closed loop.
It is equally important to switch to a manual mechanism. The system should be able to identify consecutive misunderstandings, strong emotions, high-risk keywords, and tool failures, rather than continuously guessing in order to maintain an automatic resolution rate. When transferring a call, it is necessary to pass on the confirmed information, unfinished actions, and a summary of the conversation to the human operator to avoid having the user repeat everything from the beginning. Improving the performance of long conversations does not mean that agents can delay indefinitely; the sooner uncertainties are acknowledged, the less subsequent remedial costs will be incurred.
The price cannot simply be calculated by multiplying $0.05 by the duration of the call. Back-end model inference, telephone lines, tool calls, log storage, and manual intervention all incur costs. Full-duplex communication may shorten a call, but it could also extend the conversation due to more natural interaction. When evaluating, one should consider the total cost per successful resolution, average processing time, interruption recovery rate, error operation rate, and the rate of repeated statements after transferring to manual assistance.
Language coverage also requires local testing. The same language can have different accents, speaking speeds, ways of pronouncing numbers, and customer service etiquette in different regions. Additionally, compression over telephone lines can result in the loss of sound details. The test set should include elderly people, children, non-native speakers, situations where multiple people are in the same room, and weak network conditions, rather than just having internal staff read scripts in a quiet office. Only when there are enough failed samples that reflect real-world scenarios will model upgrades not mistake demonstration results for the actual user experience.
GPT – Live – 1 is already available in API, but more voices and languages will continue to be added. However, there are also limitations on some customization capabilities. This feature allows voice agents to surpass the threshold of a “rotating reading” experience, but it does not eliminate the boundaries of responsibility for business systems. What determines whether a product is useful is not just whether the model sounds human-like, but whether users can truly stop speaking when they need to, whether background actions can be synchronously reversed, whether issues can be quickly resolved when they occur, and whether a verifiable record is left for each step of the process.
