Skip to main content
OpenAI's new GPT-Live-1.1 voice model interface, showing a waveform and cost reduction to 2 cents.

Editorial illustration for OpenAI Launches GPT-Live-1.1 Voice Model, Cuts Idle Cost to 2 Cents

OpenAI's GPT-Live-1.1 Slashes Voice API Costs to 2¢

OpenAI Launches GPT-Live-1.1 Voice Model, Cuts Idle Cost to 2 Cents

• 4 min read

OpenAI pushed GPT-Live-1.1 into its API this week under the model ID gpt-live-1.1, the successor to GPT-Live-1, which launched on September 10, 2026, and has been running ChatGPT Plus and Pro voice since October 15. The new model keeps the full-duplex trick that defined its predecessor: it listens and talks in the same stream, so a user can interrupt mid-sentence and the model adjusts without missing a beat. OpenAI says the update brings steadier accents across languages, better handling of transcription, letters and numbers, and more reliable task completion, though the company hasn't published any benchmark scores to back that up.

The bigger change for developers running voice agents is on the billing side. OpenAI added a "standby" state that charges a lower rate for idle sessions, meaning a call that's connected but not actively talking no longer burns money at the same pace as live conversation. Sessions connect over WebRTC or WebSocket through v1/live/sessions, cap out at a 128,000-token context window, and stay locked to OpenAI's own cloud, no open weights, no local deployment option.

The GPT-Live-1.1 voice model is a full-duplex model for multilingual conversations. It can listen while it speaks. OpenAI team states that it speaks with more natural and consistent accents across languages and completes tasks more reliably.

Why this matters

For anyone building voice agents, the standby pricing is the practical headline here. A two-cent-per-minute idle rate changes the math for products that hold calls open while waiting on a customer or a database lookup, which is most of them. That's a cost problem solved, not a capability leap, and it's worth treating it as such.

The full-duplex, listen-while-speaking design matters more for product feel than for benchmarks. Interruptions and overlapping speech are where voice agents fall apart in demos, so if OpenAI's claims about steadier accents and more reliable task completion hold up under real traffic, that's the piece worth testing first.

We'd also note this ships fast on the heels of GPT-Live-1, which only reached the API in September and has been running ChatGPT's voice mode since October 15. That's a tight iteration cycle for a model already in production. Worth watching whether structured outputs and fine-tuning arrive for GPT-Live-1.1 on a similarly short timeline, since that's what turns a voice demo into something teams will actually ship.

Common Questions Answered

What is the key difference between GPT-Live-1.1 and its predecessor GPT-Live-1?

GPT-Live-1.1 maintains the full-duplex capability that allows the model to listen and speak simultaneously in the same stream, enabling users to interrupt mid-sentence without the model missing a beat. The new version improves upon this foundation with steadier accents across languages and more reliable task completion, making voice interactions feel more natural and responsive.

How does the standby pricing model for GPT-Live-1.1 impact voice agent development?

The new idle cost of 2 cents per minute significantly reduces expenses for voice agents that need to hold calls open while waiting for customer responses or database lookups. This pricing change solves a major cost problem for most voice agent products, making it economically viable to maintain persistent connections during processing delays.

What multilingual improvements does GPT-Live-1.1 offer compared to previous versions?

GPT-Live-1.1 delivers more natural and consistent accents across multiple languages, addressing a key limitation in multilingual voice interactions. This improvement enhances the user experience by making the voice model sound more authentic and reliable when conducting conversations in different languages.

Why is the full-duplex listen-while-speaking design important for voice agent products?

The full-duplex design matters more for the overall product feel and user experience than for raw performance benchmarks, particularly in handling interruptions and overlapping speech naturally. This capability allows voice agents to respond to user interruptions seamlessly, creating more human-like conversations that feel less robotic and more interactive.

LIVE13:53OpenAI Launches GPT-Live-1.1 Voice Model, Cuts Idle Cost to 2 Cents