OpenAI shipped GPT-Live on July 8, 2026, a new voice model that listens and talks at the same time instead of waiting for each turn. The company says this makes ChatGPT Voice feel like talking to a real person.
What GPT-Live Actually Does
GPT-Live runs on what OpenAI calls a full-duplex architecture. Older voice AI systems worked in strict turns: you speak, the AI processes, the AI responds. GPT-Live continuously processes your input while generating its own output, making decisions many times per second about whether to speak, stay quiet, or interrupt.
The model acknowledges you mid-sentence with phrases like “mhmm” or “got it.” It pauses when you pause instead of rushing to fill silence. Background noise like traffic or nearby conversations gets filtered out better than previous versions.
Two Versions Roll Out Globally
OpenAI released two variants: GPT-Live-1 and GPT-Live-1 mini. The full version becomes the default for Go, Plus, and Pro subscribers. The mini version powers the free tier. Both are available on iOS, Android, and the web starting today.
When users ask complex questions that need web search or deeper reasoning, GPT-Live delegates to GPT-5.5 in the background while keeping the conversation flowing. Users can pick between Instant, Medium, and High reasoning modes.
Benchmarks Show Real Improvement
In head-to-head comparisons against Advanced Voice Mode, GPT-Live-1 was preferred 75.7% of the time across 5-to-10 minute conversations. On GPQA, a scientific reasoning benchmark, GPT-Live-1 High scored 84.2% compared to AVM’s 45.3%.
On BrowseComp, which tests web search capabilities, GPT-Live-1 High hit 75.2% accuracy versus AVM’s 0.7%. That’s a massive gap for any task that involves finding information online during a voice conversation.
Safety and Voice Impersonation Guards
Because GPT-Live processes audio in real time, OpenAI built safeguards that can intervene while the model is speaking. The system can redirect unsafe responses, show safety messaging, or end high-risk conversations. For self-harm scenarios, voice-adapted crisis helpline support kicks in.
Teen users get additional protections built directly into the model. GPT-Live uses predefined voices and blocks attempts to impersonate real people.
What Changed From Previous Versions
The original ChatGPT Voice chained three separate models: speech-to-text, an LLM, and text-to-speech. Information got lost between models, responses felt stilted. Advanced Voice Mode processed audio within a single model but still operated in discrete turns, with silence-based turn detection causing frequent awkward interruptions.
GPT-Live processes audio input and output continuously within a single model, and delegates heavy lifting to separate models when needed. This separation lets it use the latest frontier models for hard questions while maintaining conversational flow.
Limitations at Launch
GPT-Live does not support voice combined with video or screen sharing yet. Those features are coming soon. Legacy voice modes remain available for features that need video. For non-English languages, the model may have accent gaps or reduced fluency, which OpenAI says it is actively improving.
The API version is not available yet, but developers can sign up for early access. OpenAI plans to update the background model from GPT-5.5 to newer frontier models as they release.
Frequently Asked Questions
Is GPT-Live free to use?
GPT-Live-1 mini is the default for free ChatGPT users. GPT-Live-1 requires a paid subscription (Go, Plus, or Pro).
How is GPT-Live different from Advanced Voice Mode?
Advanced Voice Mode still uses turn-based processing where it waits for silence to respond. GPT-Live listens and speaks simultaneously, interrupting naturally and maintaining conversational flow like a real person.
Does GPT-Live work with video?
Not at launch. Voice and video/screen sharing cannot be combined in GPT-Live yet, though OpenAI says the feature is coming soon. Legacy voice modes support video.
What languages does GPT-Live support?
GPT-Live is optimized for ChatGPT’s most popular languages, but non-English speakers may encounter accent issues or fluency gaps that OpenAI is working to fix.
