What happened
OpenAI released two new conversational models this week — GPT-Live-1 and GPT-Live-1 mini. According to TechCrunch, the models are full-duplex, allowing them to speak and listen simultaneously. This enables users to interrupt the AI naturally, a key feature for live conversation and translation.
The smaller GPT-Live-1 mini will become the default Advanced Voice Mode in ChatGPT. The more capable GPT-Live-1 model will be available to users on paid tiers. The models can also present information visually.
How the room's reading it
The move is widely seen as OpenAI pushing voice to be a primary computing interface, not just a feature. The key technical shift is from a slow, three-step process — transcribe, think, speak — to a single model that can be interrupted. Developers on X note this finally solves the awkward pauses and turn-taking that plague most voice AI.
Still, some observers point to the demos, where a live Hindi translation had a heavy American accent and sounded unnatural. Competitors like Apple and Amazon are also chasing more conversational assistants, so this is OpenAI defending its position. The consensus is that while the tech is a step forward, the polish for non-English languages isn't there yet.
Sailfish's take
Everyone is focused on the user being able to interrupt the AI. We think that's looking through the wrong end of the telescope. The real unlock is that the AI can now listen while it's talking, and listen while you're talking. This is the foundation for agents that can react to non-verbal cues or a changing environment in real time.
We've built enough voice agents to know that latency and awkward turn-taking are the biggest killers of user trust. Full-duplex audio isn't just a quality-of-life improvement — it's a step-change for building anything that needs to feel present and aware. We'd start building agents that don't just wait for a command, but actively listen to a whole conversation.