GPT-Live 1.1 Launches on AI Gateway: Full-Duplex Speech Model Upgrade

GPT-Live 1.1 full-duplex speech model launches on AI Gateway with more accurate transcription, more stable voices, and low migration cost.
OpenAI's GPT-Live 1.1 is now available through AI Gateway, with full-duplex voice interaction as its core feature: the model keeps listening while it speaks, so users can pause, interrupt, or add information at any time, making the interaction rhythm feel close to a real conversation. Compared with the previous generation, the new version improves speech transcription accuracy and maintains more consistent accents and voice characteristics across languages. For existing integrations, upgrading is nearly cost-free—just replace the model ID with openai/gpt-live-1.1; the endpoint and client delegation mechanism remain the same. Client delegation allows complex tasks encountered in a voice session to be handled by any text model on AI Gateway, with the result announced via voice, and delegated requests are billed separately. AI Gateway itself provides unified invocation, usage tracking, retries, and failover, syncs with provider pricing without markup.
OpenAI's GPT-Live 1.1 is now available through AI Gateway. As a full-duplex speech model, its standout feature is "listening while speaking"—callers can pause, interrupt, or add details at any point while the model is replying, making the interaction rhythm feel much closer to a real human conversation. Compared with the previous generation of GPT-Live, the new version improves speech transcription accuracy and maintains more consistent accents and voice characteristics across languages.

What Full-Duplex Speech Means
Traditional speech models mostly use a "turn-taking" mode: the user finishes speaking, then the model responds. GPT-Live 1.1's full-duplex design lets the model keep listening while it speaks, which means users can interrupt it the way they would interrupt a real person, or add information halfway through its reply. For scenarios like voice assistants, customer service bots, and real-time translation, the improvement in naturalness is obvious.
Full-duplex is originally a communications term referring to both parties being able to transmit data in both directions simultaneously—a typical example is a telephone call, whereas a half-duplex walkie-talkie requires one party to finish speaking and press a button to switch direction. In speech AI, the "turn-based" mode is equivalent to half-duplex: the model must wait for the user's speech to end (usually relying on voice activity detection, VAD, to determine silence) before it starts inferring and synthesizing a reply, so end-to-end latency inherently includes the time spent waiting for the user to finish. A full-duplex model continuously receives the audio stream and continuously generates output, maintaining listening and comprehension while the user speaks, thereby supporting real-time interactions such as barge-in. The trade-off is that the model must handle both "listening" and "speaking" signals simultaneously, which places higher demands on inference architecture and latency budgets, and makes it harder to handle semantic alignment when user speech and model output overlap.
Upgrade Highlights: More Accurate Transcription, More Stable Voice
According to the official documentation, GPT-Live 1.1 brings clear improvements in two areas:
- Higher speech transcription accuracy: The model recognizes user speech more reliably, reducing irrelevant answers caused by mishearing.
- Better cross-language voice consistency: When switching between multiple languages, the model maintains more consistent accents and voice characteristics, avoiding the jarring feeling of "a different person for a different language."
These two points are especially important for products that require long voice interactions—users are very sensitive to voice continuity and recognition accuracy.
Low Migration Cost: Just Change the Model ID
For developers with existing GPT-Live integrations, upgrading to 1.1 is nearly cost-free. The new model uses the same Live endpoint and client delegation mechanism; you only need to replace the model ID with openai/gpt-live-1.1. In the official AI SDK example, you specify the model and the api: 'live' parameter through the experimental_realtime method, then retrieve the WebSocket configuration to establish a session.
Client Delegation: Calling Text Models Within a Voice Session
Another capability of GPT-Live 1.1 is client delegation. When enabled, developers can tell the model in the session instructions when it needs external help. When the model encounters a problem requiring additional processing (such as an order lookup), it sends a delegation-created event; upon receiving it, the application can use any text model on AI Gateway to handle the request, then return the result to GPT-Live through the commentary channel, which announces it via voice.
The official example uses openai/gpt-6.2-sol to complete the delegation task. The application retains control over the delegated work, including permissions, confirmation, and cancellation. Note that delegated model requests are billed separately through AI Gateway, and usage of the voice session continues to be counted while the delegation is being executed.
The "client" in "client delegation" refers to the application calling GPT-Live, not the end user's browser. The speech model itself excels at low-latency listening and speaking interactions, but has limited capability for tasks requiring database queries, business API calls, or multi-step reasoning; the delegation mechanism "outsources" such subtasks. The model notifies the application via an event that it needs help, and the application decides which text model to call, what context to include, and whether to require user confirmation, then passes the structured result back. This design keeps tool invocation and permission control on the application side; the speech model is only responsible for deciding "when help is needed" and "how to express the result in speech," and it also lets developers flexibly choose text models at different tiers based on task cost, without having to cram all logic into the real-time voice session.
AI Gateway Platform Capabilities
AI Gateway itself provides a unified model invocation API, supports usage and cost tracking, and is configured with retries, failover, and performance optimization to deliver higher service availability than any single provider. The platform also includes custom reports, zero data retention support, API key budgets, routing rules, and more. On pricing, AI Gateway syncs with provider prices without markup and charges no platform fee for inference; BYOK (bring your own key) requests are treated the same way.
BYOK (Bring Your Own Key) means developers use their own API keys, obtained and paid for directly with the model provider, with requests forwarded through the gateway. In this mode, inference costs are settled directly with the provider, and the gateway typically only provides routing, observability, and failover capabilities, so not charging an additional inference platform fee is a reasonable business model; by contrast, using gateway-managed keys may involve top-ups or quota management. Failover means that when the primary provider returns an error, times out, or triggers a rate limit, the gateway automatically retries the request to a backup provider or backup model, thereby improving overall availability. Zero Data Retention commits that request and response content is not persistently stored, which is often a hard requirement in compliance assessments for products involving user voice, medical, or financial information.
Getting Started
Developers can try GPT-Live 1.1 directly in the model playground; the complete integration guide (including WebSocket connection, audio stream transmission, and full examples) is available in the GPT-Live guide. The official sample code omits connection establishment, event parsing, transcript assembly, audio streaming, and closing; these parts need to be filled in for actual integration.
For teams building real-time voice interaction products, GPT-Live 1.1's full-duplex capabilities and low migration cost are worth attention; and the client delegation mechanism provides a clear implementation path for a hybrid "speech model + text model" architecture.
Related articles

Gemini 4 Neon Hands-On Test: Can It Replace Opus 5.5? Three-Round Comparison Reveals the Answer
A Bilibili creator's three-round hands-on test compares Gemini 4 Neon, Opus 5.5, and GPT-6.2 Sol across 3D modeling and animation, analyzing performance and pricing to see if Gemini 4 Neon can replace Opus 5.5.

Claude Fable 5.5 Leaked, GPT-6.2 Launch Sparks Controversy as AI Model Wars Escalate
Anthropic's mysterious Silicon 2 may be Claude Fable 5.5, OpenAI's GPT-6.2 Astra and Soul spark polarized community reactions, while Google Gemini Neon and task features plus Microsoft Ignite models arrive. This article sorts out each model's real strengths and weaknesses and the industry signals.

Real-Life Jewel Thief on GTA 6 Heist Mechanics: Casing Is the Core
Former jewel thief Larry Lawton breaks down real heists on Lex Fridman's podcast—casing and reading people's habits—and shares what he expects from GTA 6's heist mechanics.