OpenAI’s Voice Agents Get Cheaper Standby Sessions and Broader Reach

Voice agents just gained a cheaper way to wait. OpenAI has made GPT-Live-1.1 available through its API under the model ID gpt-live-1.1, adding a standby billing state that cuts the cost of silent sessions to $0.02 per minute.
That change targets teams running voice agents across long calls, support workflows, and phone interactions. The model can listen while it speaks, handle multilingual conversations, and keep an open WebRTC connection while pausing inference during silence.
A Voice Model Built for Continuous Conversation
GPT-Live-1.1 is a full-duplex speech model from OpenAI that hears and speaks in 1 stream. Its inputs and outputs are audio and text, and it can manage turn detection while returning ASR transcripts natively. Streaming and function calling work, but image and video inputs are not supported.
The GPT-Live family replaces the usual cascade of speech-to-text, an LLM, and text-to-speech. OpenAI says GPT-Live reasons over incoming and outgoing audio together in a single model, allowing it to respond inside the flow of a conversation instead of waiting for each separate stage to finish.
OpenAI lists four improvement areas for GPT-Live-1.1: accents, transcription, letter and number handling, and task completion. The company says the model speaks with more natural and consistent accents across languages and completes tasks more reliably than GPT-Live-1.
There are no published benchmark scores for the new model. OpenAI published 0 benchmark scores for 1.1, so the product claims are not paired with a public scorecard.
GPT-Live-1.1 follows GPT-Live-1, which reached the API on September 10, 2026. The new model has powered ChatGPT Plus and Pro voice since October 15, and outbound SIP calling reached all organizations on November 2, so phone agents are supported.
Standby Billing Turns Silence Into a Smaller Expense
The headline change for developers is standby. When a session becomes quiet, standby pauses inference while the WebRTC connection stays open, allowing the conversation to resume without throwing away the session.
- Active: $0.05 per minute, or $3.00 per hour
- Standby with WebRTC and inference paused: $0.02 per minute, or $1.20 per hour
Voice sessions are billed by duration, and the GPT-Live-1 model page says billing is calculated per second. The standby rate is 60% lower than the active rate, giving teams a direct way to reduce idle call costs. Backend model and tool usage are billed separately.
The Live conversations guide documents three standby settings: session.wake_policy.type, session.idle_timeout_s, and session.wake_policy.vad_threshold. After the timeout, the server emits session.paused. When the user speaks, session.resumed fires, while the session ID and conversation history remain available.
OpenAI warns that the first reply after a short utterance may be slower. Developers must keep the microphone track active during standby, and the feature works with both GPT-Live-1 and GPT-Live-1.1.
Live sessions hold a 128,000-token default context window, while instructions can run up to 16,384 tokens. The model runs only on OpenAI’s cloud through v1/live/sessions, and clients connect through WebRTC or WebSocket. There are no open weights or quantized variants.
Delegation, Dictation, and Payments Expand the Voice Push
GPT-Live-1.1 can hand deeper reasoning and tool calls to a backend text model. OpenAI names GPT-6 Astra for complex issues and Luna for high-volume tasks, while third-party backends also work.
The Live conversations guide sets delegation.type to either client or responses. Responses mode uses OpenAI models; client mode connects a team’s own backend. Developers can switch an existing Live session by setting the model to gpt-live-1.1, but changing the model requires a new session.
The model is API-only, and GPT-Live-1 and GPT-Live-1.1 have separate concurrent-session limits. Each limit applies across inference regions, but the GPT-Live-1.1 page does not publish tier numbers. Structured outputs, fine-tuning, and predicted outputs do not work.
OpenAI is also pushing voice interaction inside ChatGPT. Live Dictation transcribes audio prompts as users speak instead of waiting until recording ends, and it removes the previous 10-minute recording limit. Users can tap or click the audio waveform to view the live transcript, then edit it in the prompt box.
The feature is rolling out globally on all ChatGPT plans across the web, desktop app, and iOS and Android apps. OpenAI says Live Dictation is getting “more reliable,” following an overhaul of ChatGPT’s two-way conversational voice mode earlier this year. ChatGPT’s prompt box dictation can also transcribe voice messages or turn them into bullet points, while record mode remains in beta for Plus, Enterprise, Edu, Business, and Pro workspaces.
Last week, OpenAI launched ChatGPT Wallet, its second attempt at agentic payments after shelving in-chat checkout earlier this year. Agents can pay merchants on a user’s behalf, but every purchase still requires human approval. OpenAI is not taking a cut of Wallet transactions at launch, while delivery and returns remain with merchants and existing card dispute protections stay in place.
That human approval requirement highlights the larger challenge for agentic payments. Google’s Agent Payments Protocol uses cryptographically signed mandates to record what a user asked for and approved, but useful agent permissions must also capture priorities, accepted trade-offs, rejected trade-offs, and the conditions that require escalation.
A spending limit tells an agent how much it can spend; a mandate tells it what the money is for. Clear liability, a duty of loyalty, and rules for conflicts of interest will decide whether consumers trust agents with wallets. The companies that build those protections into agentic systems now will shape whether these systems scale or stall after a visible failure.
GPT-Live-1.1 makes the voice side of that future more practical by joining continuous speech, backend delegation, phone support, and lower-cost waiting into one API path. The next test is no longer whether agents can talk. It is whether they can act with enough reliability, control, and accountability for people to keep listening.
Based on
- OpenAI Ships GPT-Live-1.1 for Voice Agents, Cuts Idle Cost to $0.02/Min — marktechpost.com
- ChatGPT can transcribe as you speak now with Live Dictation | The Verge — theverge.com
- Openai Archive for November 2026 – Page 1 | The Verge — theverge.com
- Your AI agent has a wallet. Who answers for its decisions? | Fortune — fortune.com




