OpenAI launches GPT-Live-1 API: full-duplex voice, telephony, and $0.05/min pricing

4 hours ago

5 min read

OpenAI has launched GPT-Live-1 in the API, extending its full-duplex voice technology to developers building customer-service systems, phone agents, assistants, tutoring products, and other real-time voice applications.

GPT-Live-1 can listen while it is speaking, instead of forcing every interaction into a rigid sequence of user turn followed by assistant turn.

The model can react to interruptions, acknowledgements, pauses, changes of direction, and background speech while audio is still flowing.

OpenAI prices the GPT-Live-1 voice layer at $0.05 per minute, although reasoning performed by backend models and external tools can add separate costs.

The resulting architecture separates the conversational voice layer from the systems performing deeper reasoning, retrieval, and actions.

··········

GPT-LIVE-1 USES A FULL-DUPLEX VOICE ARCHITECTURE.

The model can process incoming speech while generating outgoing audio, reducing the dependence on sequential speech-to-text, language-model, and text-to-speech stages.

Traditional production voice agents often rely on a pipeline such as speech-to-text → language model → text-to-speech.

Each stage adds latency and can remove information about timing, hesitation, interruptions, and conversational rhythm.

A short pause can be mistaken for the end of a turn, while an interruption can arrive after the system has already committed to a response.

GPT-Live-1 keeps incoming and outgoing audio active at the same time.

The model can decide continuously whether to keep listening, continue speaking, pause, respond to an interruption, or pass work to another system.

Developers can still use turn detection when explicit conversational boundaries are useful.

The API also exposes ASR transcripts and response text, allowing voice applications to retain textual records of both user input and model output.

........

GPT-Live-1 API architecture at launch.

Capability	GPT-Live-1
Interaction architecture	Full duplex
Simultaneous listening and speaking	Yes
Interruptions and backchannels	Native handling
Speech transcript output	Yes
Response text output	Yes
Turn detection	Supported
Backend reasoning delegation	Supported
Third-party backend model delegation	Supported
Telephony use	Supported
Voice-layer price	$0.05 per minute
Custom voices	Restricted access / eligibility process

........

GPT-Live-1 does not need to perform every reasoning task itself.

It can maintain the live conversation while another model or external tool handles work that requires more compute, data access, or longer reasoning.

In production terms, the voice model becomes the real-time interaction layer of a broader agent stack.

··········

GPT-LIVE-1 CAN DELEGATE REASONING TO GPT-6 ASTRA OR OTHER MODELS.

Developers can keep the voice interface responsive while assigning different backend models according to task complexity, latency requirements, and cost.

OpenAI describes configurations in which GPT-Live-1 manages the conversation while models such as Luna process high-volume or simpler requests and GPT-6 Astra handles more demanding cases.

Third-party models can also operate behind the voice layer.

A voice system therefore does not need to run its most expensive reasoning model continuously simply because a call remains active.

A basic scheduling request could be routed to a lower-cost model, while a complicated account dispute could trigger a more capable reasoning backend.

The same design applies to tools.

GPT-Live-1 can continue managing the call while another component searches a database, accesses an application, retrieves account information, or executes an agentic workflow.

OpenAI also reports improved handling of silence and background noise, reducing false interruptions caused by normal pauses or environmental sound.

Developers can control tone, pacing, and conversational behavior through system instructions, instead of treating the generated voice as a separate fixed output stage.

The API also supports keyword biasing and stronger alphanumeric recognition, useful for names, reference codes, order numbers, addresses, account identifiers, and product codes.

Telephony support makes GPT-Live-1 directly applicable to call-center and phone-agent systems.

Likely workloads include customer support, reservations, appointment booking, collections, sales qualification, internal help desks, and other processes where the model has to speak and complete actions during the same interaction.

··········

$0.05 PER MINUTE IS THE VOICE-LAYER PRICE, NOT THE FULL AGENT COST.

The headline rate covers GPT-Live-1 itself, while delegated inference, tool calls, retrieval, databases, and surrounding infrastructure can increase the cost of each completed interaction.

At $0.05 per minute, the raw voice-layer cost scales directly with conversation duration.

That makes basic capacity planning straightforward.

........

Data Studios calculation: GPT-Live-1 voice-layer cost at different usage levels.

Monthly GPT-Live-1 time	Voice-layer cost
1,000 minutes	$50
10,000 minutes	$500
100,000 minutes	$5,000
500,000 minutes	$25,000
1,000,000 minutes	$50,000

........

A company processing 100,000 minutes of GPT-Live-1 conversations per month would spend about $5,000 on the voice layer alone.

With an average session length of five minutes, 100,000 minutes would represent approximately 20,000 five-minute conversations, equivalent to around $0.25 in GPT-Live-1 voice cost per session.

The complete cost of those conversations can be higher.

Calls to GPT-6 Astra or another backend model, searches, database operations, tool execution, and long reasoning sequences all sit outside the headline $0.05-per-minute price.

For cost comparisons, the voice rate alone is therefore insufficient.

A more useful production metric is cost per resolved call or completed task.

An agent with a higher per-minute cost can still be cheaper overall if it finishes conversations faster, resolves more cases without human escalation, or reduces repeat calls.

OpenAI has also cited an early implementation in which replacing a cascaded voice stack with GPT-Live-1 reduced the relevant codebase by around 80% and removed approximately 23,000 lines of code.

That figure comes from a specific customer deployment and should not be generalized to every voice system, but it shows that engineering complexity can materially affect the economics alongside inference pricing.

··········

THE FIRST GPT-LIVE-1 BENCHMARKS MEASURE CONVERSATIONAL BEHAVIOR.

OpenAI reports large gains on full-duplex interaction tests, while end-to-end agent quality still depends on the backend model, tools, and production environment.

OpenAI reports a 30-percentage-point improvement on Full Duplex Bench compared with GPT-Realtime-2.1.

The benchmark focuses on behavior specific to live speech, including interruptions, turn taking, background speech, and listener backchannels.

Conventional reasoning benchmarks do not measure whether a voice agent interrupts users, waits too long before answering, or mishandles a correction spoken in the middle of a response.

OpenAI also reports that GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranks first on Tau3, an evaluation focused on end-to-end voice-agent task performance.

These are launch-time vendor results.

Independent production testing will still need to measure latency distribution, telephony reliability, accent handling, noisy environments, long-session stability, tool-call success, backend inference cost, and completed-task rates.

OpenAI has also cited testing from language-learning company Speak, which reported that GPT-Live-1 reduced interruptions during learner thinking pauses by almost 80% compared with its previous turn-based systems.

The available real-time voice selection is also expanding across additional accents, dialects, and languages.

Custom voice access remains subject to a separate eligibility process rather than being universally available through the standard API.

GPT-Live-1 effectively turns voice into an agent interface rather than a speech layer attached to a text chatbot.

The live model manages timing and conversation, backend models handle reasoning, and tools perform external actions.

For production systems, the key measure will increasingly be whether the agent can resolve the task during the call at an acceptable total cost, not simply whether its synthetic voice sounds natural.

··········

·····

DATA STUDIOS

·····

datastudios.org