Skip to main content
What's new. Why it matters.
AI & Tech•Model & Runtime Release

OpenAI Opens GPT-Live-1.1 to Developers With $0.02/Min WebRTC Standby Mode, 38 Voices, and Backend Agent Delegation

Following its October rollout in ChatGPT Voice, OpenAI’s full-duplex GPT-Live-1.1 voice model is now live in the OpenAI API—introducing opt-in WebRTC standby billing that slashes idle session costs by 60%, sharper multilingual accent consistency, and decoupled backend reasoning delegation.

By TrendyFlash Editorial Desk (Editorial Desk)
Published: • 7 min read
Tight 85mm close-up documentary photograph of an audio systems engineer’s hands adjusting a brushed-aluminum rotary gain dial on a matte-black broadcast USB-C audio interface beside a studio condenser microphone boom arm and subtle emerald-amber LED level meter
OpenAI’s November 19 API release of GPT-Live-1.1 brings full-duplex conversational audio, 60%-discounted WebRTC standby sessions, and 38 voices to production voice agents.Credit: TrendyFlash Editorial Photography

At a Glance

  • What changed: OpenAI opened its upgraded full-duplex voice model, gpt-live-1.1, to developers via the OpenAI API on November 19, 2026—expanding its October 15 ChatGPT Voice rollout with opt-in WebRTC standby mode, 38 voice IDs (including new Spanish-leaning voices), improved multilingual transcription and accent stability, and backend model/tool delegation.
  • Who can use it: Developers building real-time voice agents, contact-center bots, and multimodal apps can invoke gpt-live-1.1 immediately over WebRTC via the OpenAI API, while consumer ChatGPT Plus, Pro, Codex, and Dots users already have access to GPT-Live-1.1 (and its lightweight GPT-Live-1.1 Quip sibling).
  • What it costs: Active full-duplex voice inference is billed per second at $0.05 per minute, while opt-in WebRTC standby drops idle waiting time by 60% to $0.02 per minute (with configurable 5–60 second idle timeouts). Delegated backend frontier models and tool calls are billed separately.
  • Why should I care: Real-time voice apps have historically burned compute budgets during natural pauses, hold queues, and reading silences; pairing $0.02/min standby with decoupled backend reasoning finally makes always-on enterprise voice interfaces economically viable at scale.

OpenAI has officially opened GPT-Live-1.1 (gpt-live-1.1) to developers via the OpenAI API on November 19, 2026—bringing its upgraded full-duplex voice architecture, 38 selectable voices, and a brand-new 60%-discounted WebRTC standby pricing tier out of consumer ChatGPT apps and into production engineering stacks. The developer release follows GPT-Live-1.1’s October 15 debut as the default real-time voice engine across ChatGPT Voice, Codex, and Dots (where it shipped alongside a lighter sibling, GPT-Live-1.1 Quip).

The Unit-Economics Breakthrough: $0.05/Min Active vs. $0.02/Min WebRTC Standby

For engineering leads and FinOps teams evaluating real-time audio infrastructure (Model Pricing & Inference), the headline feature of the gpt-live-1.1 API rollout is its explicit, opt-in WebRTC standby mode. Previously, keeping a low-latency voice socket open while a user paused to read a screen, verify a two-factor code, or talk to someone in the room forced developers to choose between burning full inference rates on silence or tearing down and re-establishing the WebRTC handshake.

Under the new per-second billing structure, active full-duplex inference costs $0.05 per minute, while idle WebRTC standby time drops 60% to $0.02 per minute. Crucially, standby is not an automatic silence rebate—developers must opt in via session configuration by setting session.wake_policy.type to speech and tuning session.idle_timeout_s to an integer between 5 and 60 seconds. Once no input or output transcript occurs across that window, the WebRTC connection stays alive while active inference pauses, waking automatically on the next spoken utterance (with a modest wake-up latency trade-off on short opening phrases). Simply muting a microphone will not trigger standby early until the configured idle timer elapses.

Capability / Pricing Dimension gpt-live-1.1 Active Mode gpt-live-1.1 WebRTC Standby Mode
API Rate (Per-Second Billing) $0.05 / minute $0.02 / minute (60% reduction)
Activation & Session State Default on connection; continuous full-duplex listening & speaking Opt-in via session.wake_policy.type = "speech" & session.idle_timeout_s (5–60s)
WebRTC Connection Lifecycle Active media stream + live speech-to-speech inference WebRTC peer connection stays open; inference pauses until speech resumes
Voice Catalog & Languages 38 voice IDs (incl. calandria, hornero, luciernaga); upgraded FR, DE, HI, JA, KO, ES-MX Preserves active session voice ID, persona, and conversation state across wakes
Backend Reasoning & Tools Delegated backend frontier models & tool calls billed separately No backend model tokens consumed while session remains idle in standby

Full-Duplex Pacing, Multilingual Accents, and 38 Voices

Architecturally, gpt-live-1.1 preserves the native full-duplex foundation of GPT-Live-1—meaning the model listens and speaks simultaneously rather than relying on brittle turn-taking voice-activity detectors. Callers can interrupt mid-sentence, add clarifying context, or steer responses in real time without jarring cutoffs. Compared with version 1.0, OpenAI’s 1.1 update targets the sharpest pain points reported by global deployments: transcription accuracy, pronunciation fidelity, and accent drift across non-English conversations.

Specifically, OpenAI highlights stronger voice and accent stability across French, German, Hindi, Japanese, Korean, and Mexican Spanish, alongside an expanded API catalog of 38 distinct voice IDs—incorporating new Spanish-leaning voices such as calandria, hornero, and luciernaga, plus broader English selections introduced in the consumer rollout. The model also expands its repertoire of natural conversational cues and brief filler acknowledgments (such as short thinking or search transitions) so callers know when a request is actively processing rather than hanging in dead air.

Decoupled Backend Delegation for Production Voice Agents

Beyond raw audio fidelity, the November 19 API launch underscores a broader shift in how enterprise teams architect conversational AI (Cloud & Developer Infrastructure). Rather than forcing a single voice model to juggle low-latency prosody, complex SQL lookups, and multi-step policy reasoning simultaneously, gpt-live-1.1 supports backend delegation—letting the real-time voice layer manage natural dialogue and interruptions while independent frontier models or backend agents execute web searches, database queries, and tool calls asynchronously.

Because those delegated backend model tokens and tool invocations are billed separately from the $0.05/$0.02 per-minute voice layer, architects can route routine dialogue through gpt-live-1.1 and invoke heavier reasoning models only when a customer’s request genuinely demands it. Combined with configurable WebRTC standby timeouts, OpenAI’s latest release gives contact-center platforms, field-service assistants, and voice-native developer tools a much tighter handle on both latency and monthly cloud spend.

Why It Matters

Voice agents have crossed the latency threshold over the past year, yet unit economics remained the silent killer of production deployments: keeping a live speech-to-speech session open while a customer reads an order number, navigates a web form, or sits on brief hold meant paying full active inference rates for dead air. By pairing full-duplex interruption handling in gpt-live-1.1 with an explicit, developer-controlled WebRTC standby state at $0.02 per minute—and separating fast conversational audio from heavier backend reasoning models—OpenAI is turning real-time voice from a flashy demo into predictable, line-item-auditable cloud infrastructure.

What Is Confirmed

Confirmed across primary OpenAI documentation and official release announcements: (1) OpenAI initially rolled out GPT-Live-1.1 (alongside lightweight GPT-Live-1.1 Quip) as the default voice engine for ChatGPT Plus, Pro, and higher-tier users across ChatGPT Voice, Codex, and Dots starting October 15, 2026; (2) on November 19, 2026, OpenAI launched model ID gpt-live-1.1 in the OpenAI API with full-duplex listening and speaking, 38 selectable voice IDs (including new Spanish-leaning voices such as calandria, hornero, and luciernaga), and improved transcription accuracy and accent consistency across French, German, Hindi, Japanese, Korean, and Mexican Spanish; (3) per-second API billing is set at $0.05/minute for active voice inference and $0.02/minute for opt-in WebRTC standby mode, configured via session.wake_policy.type set to speech and session.idle_timeout_s between 5 and 60 seconds; and (4) muting a client microphone alone does not trigger standby until the idle timeout expires, while any delegated backend reasoning models or tool/function calls are billed separately from the voice session layer.

What Remains Unclear

OpenAI has not yet published standardized third-party latency benchmarks comparing wake-from-standby first-utterance overhead against continuously active sessions, nor has it detailed whether the lightweight GPT-Live-1.1 Quip consumer variant will receive its own dedicated lower-cost API endpoint or volume-commitment tiers for enterprise telephony carriers.

Sources & Underlying Evidence (2)

We link directly to the primary documents supporting our reporting and distinguish company claims from independently checked facts.

More in AI & Tech

Two AI infrastructure engineers walking outdoors along a sunlit brutalist limestone campus terrace balustrade under crisp morning daylight while reviewing printed latency benchmarks and a slate tablet
AI & Tech•Model & Runtime Release

Google Launches Gemini 4 Neon and Stitch Loop as Cloudflare Open-Sources Clairic and Vercel Debuts Multi-Agent Board

Production AI infrastructure pivoted toward lower-cost steerability, specialized extraction, and swarm coordination on November 18 as Google rolled out $3/M Gemini 4 Neon alongside 24/7 Stitch Loop agents, Cloudflare’s Workers AI Superintelligence neolab released open-weight Clairic and Clairic-flash with zero output fees, and Vercel shipped Board for multi-agent software factories.

8 min read