Voice conversations with AI systems have long been marked by a familiar pattern: brief, awkward pauses filled with repetitive filler phrases while a response is prepared.
To make ChatGPT Voice both faster and more natural, OpenAI is upgrading the feature's underlying architecture with its new 'GPT-Live-1.1' model and a lighter, faster variant named 'GPT-Live-1.1 Quip.'
The update rolls out across ChatGPT, Codex, and dots, bringing new voices, improved regional accents, and a significant reduction in the repeated "checking" phrases that previously interrupted long-running tasks. Instead of repeatedly chiming in with variations of "Let me check on that" while waiting for search results or processing background requests, the model now manages silent working periods and delegated tasks far more naturally.
This release builds directly upon the original GPT-Live model family introduced in July.
Built with full-duplex capabilities, the system can listen and speak simultaneously rather than waiting for a rigid end to a user's turn, allowing it to seamlessly adapt to pauses, interruptions, and sudden shifts in conversational pace.
Under the hood, the architecture preserves a two-model setup: one layer maintains the continuous spoken exchange while handing off heavier reasoning or tool use to a separate backend model.
When API access launched in September, the voice layer was priced at five cents per minute billed by the second, with backend model execution billed separately. Under the new 1.1 framework, GPT-Live-1.1 becomes the default for Plus and higher-tier accounts across ChatGPT, Codex, and dots, while Quip serves as the default for standard users.
Beyond workflow management, the update significantly expands the assistant's vocal versatility.
OpenAI has added 13 new voices across the platform, introducing richer localized accents for French, German, Hindi, Japanese, Korean, and Mexican Spanish, alongside a new English voice called "Bubbie."
Users can select these new audio profiles in Settings under Voice, though specific voice availability varies depending on regional rollouts and account tiers.
The upgrade also provides smoother multilingual handling and enhanced steerability, making it effortless to guide the assistant mid-sentence or switch between languages without losing context.
When the model needs to query the web, it now employs a much broader and more varied selection of short phrases rather than relying on a predictable script.
Safety and alignment controls for the 1.1 update follow the established framework of original training, voice-specific evaluations, and ongoing session monitoring.
According to the system card released alongside the update, production evaluations using real user audio demonstrated performance at full parity with predecessor models. However, synthetic edge-case testing highlighted minor regressions in specific safety categories for the main model, including sexual content, mental health, emotional reliance, and gore, which OpenAI noted as key areas for ongoing refinement.
While developers await the upcoming API rollout for the 1.1 models, everyday users are already experiencing a more fluid communication loop that replaces rigid, walkie-talkie style exchanges with a genuine and responsive dialogue
GPT-Live itself arrived in July as a full-duplex model family that can listen and speak at the same time, rather than waiting for a clear end to a user's turn. It can follow pauses, interruptions, and shifts in pace, and it hands more complex reasoning or tool use to a separate backend model while the spoken conversation continues.
The original versions became the default for ChatGPT Voice, with the fuller model for higher-tier accounts and a mini variant for free users. The September API release priced the voice layer at five cents per minute, billed by the second, with backend model usage charged separately.
A system card published alongside the 1.1 update describes the models as continuing the full-duplex design, with GPT-Live-1.1 serving as the default for Plus and higher tiers (including Codex and Dots) and Quip as the default for other users.
Safeguards follow the same general approach as prior versions, drawing on training, voice-specific evaluations, and existing monitoring.
Production evaluations using real user audio showed performance at parity with predecessors, while some synthetic edge-case tests recorded regressions in categories such as sexual content, mental health, emotional reliance, and gore for the main model, areas the company noted for further work.
The changes leave the underlying two-model structure in place: one layer manages the continuous spoken exchange, and another handles delegated work.
For users, the immediate differences appear in the range of available voices, the accents those voices use, and the variety of brief spoken signals that replace a narrower set of filler phrases during searches or tool calls. The rollout applies across the existing Voice surfaces in ChatGPT, Codex, and dots, with the same permission and account controls that already govern those features.
Further reading: OpenAI Puts 'GPT-Live-1' in the API as Voice Agents Learn to Talk Over Themselves
