The editorial argues that 232ms median latency puts GPT-Live inside the ~300ms window humans use to judge conversational turn-taking. Crossing this threshold is framed as a qualitative shift — conversation stops feeling like walkie-talkie exchanges and starts feeling like being in a room with someone.
The editorial highlights that tool calls now execute in parallel with continuing speech synthesis, eliminating the dead-air pauses that plagued the Realtime preview. This solves the long-standing three-way tradeoff between latency, tool use, and interruption handling that every prior voice API has struggled with.
By submitting the OpenAI launch post and driving it to 676 points with 443 comments, the submitter and voting community signaled this is a notable milestone in the voice AI space. The strong engagement suggests developers see real-time multimodal streaming as a category-defining capability.
OpenAI shipped GPT-Live today — a persistent bidirectional streaming API that fuses audio, video, and text into a single session. The headline number: median voice round-trip latency of 232ms, with a 95th percentile under 400ms. That's not just faster than the old Realtime API preview (which routinely hit 800ms tail latencies); it's inside the roughly 300ms window humans use to decide someone has stopped talking. Cross that threshold and conversation stops feeling like radio and starts feeling like a room.
The product is a WebSocket (or WebRTC) session that stays open for up to 30 minutes per connection. You stream in Opus audio, H.264 or raw frames for vision, and text events. It streams back audio, transcripts, tool-call events, and — new here — mid-turn visual annotations for cases where the model is watching a screen share or camera feed and needs to point at something. Pricing is $0.06 per session-minute for audio-only, $0.18 with vision. Tokens still get billed for tool call payloads, but the main meter is wall-clock.
Under the hood, OpenAI confirms GPT-Live runs a distilled variant of GPT-5o they're calling 5o-flash-live, tuned for streaming inference with speculative decoding on the audio decoder. They also disclosed something more interesting: tool calls now execute in a side-channel while the audio stream keeps flowing. In the Realtime preview, calling a function paused speech synthesis; in GPT-Live, the model can keep saying "let me check that for you" while it actually checks.
Every voice API before this — OpenAI's own Realtime preview, Google's Gemini Live, ElevenLabs' Conversational AI, Retell — has been fighting the same three-way tradeoff: latency, tool use, and interruption handling. You could have two. If you wanted low latency you skipped tools. If you wanted tools you paid with a dead-air pause. If you wanted clean interruptions you fought VAD tuning for weeks. GPT-Live is the first API where you don't feel the tradeoff in the first ten minutes of building on it.
The pricing shift is the underrated story. Moving from per-token to per-session-second billing changes which products are economically viable overnight. A customer support agent that talks for six minutes now costs about $0.36 in audio — cheaper than a per-token equivalent once you factor system prompts and RAG context that used to get re-billed on every turn. But a companion app where the user leaves the mic on all day is now catastrophically expensive. Expect a wave of apps that aggressively close and reopen sessions on silence, and a corresponding wave of complaints when reconnection introduces the very latency the API was built to eliminate.
The community reaction on Hacker News is unusually pointed. The top comment (676 points at the time of writing) is a working demo from a solo dev who wired GPT-Live into a home robotics arm and got it to hand him a screwdriver in about a second — including the tool-call round-trip to a local ROS node. That's the shape of the demo that actually matters here: the tool call happened while the model was still describing what it was doing, so the audio never stalled. Every prior 'AI voice agent' demo had a suspicious two-second pause where the model was 'thinking'; GPT-Live removes the pause and reveals how much the pause was doing to break the illusion.
Compare against Google's Gemini Live, which shipped a similar API last quarter. Gemini Live's latency is competitive on paper (around 280ms median) but degrades sharply when you add tool calls — and its pricing is still per-token, which means a chatty session with heavy context can silently 3x your bill. Anthropic hasn't shipped a native streaming voice API at all; their bet has been that agentic tool use with the standard messages API is enough, which is starting to look like the wrong bet if the next generation of consumer AI products is voice-first.
One worth flagging: the vision channel. GPT-Live accepts continuous video at up to 4fps and returns bounding box annotations. This is not a general-purpose vision API — it's tuned for "watch a screen share and help the user" or "watch a camera and describe what changes." The frame budget is small, and the model will absolutely miss things that flicker. But for the narrow use case of a live copilot watching your workflow, it works, and it works cheaply.
If you're building voice agents: rip out your VAD tuning code. GPT-Live handles interruption detection server-side and it's better than what you'd build. The main integration work now shifts to your tool layer — specifically, making sure every tool your agent can call responds in under 500ms, because anything slower and the model starts filling with obvious stalling phrases ("let me pull that up for you") that users learn to hate. The bottleneck in voice AI moved from the model to your backend, and most backends aren't ready for it.
If you're building agentic workflows: the mid-stream tool call is quietly a big deal for non-voice use cases too. You can now build agents that narrate their reasoning while executing steps, which is either a UX win or a disaster depending on how you frame it. For debugging and observability, watching an agent think out loud in real time is genuinely useful; for end users, it can feel like being trapped in someone else's stream of consciousness.
If you're pricing this into a product: model your session length distribution carefully. The median won't hurt you; the tail will. Any user who forgets to close the session, or any app that keeps sessions open through natural silences, will burn budget fast. Build in aggressive idle timeouts (30-60s of silence should trigger a graceful teardown) and cheap session resumption via context checkpointing. OpenAI charges a small reconnection fee but it's much less than idle session-seconds.
If you're skeptical: the 232ms median is measured from OpenAI's own infrastructure to test clients in the same region. Real-world numbers with a mobile client on flaky wifi will be worse, sometimes much worse. WebRTC helps but doesn't magic away network reality. Test in your actual deployment environment before promising anything to stakeholders.
The interesting bet OpenAI is making is that the next unlock in AI products isn't intelligence — it's presence. GPT-5's benchmarks are impressive but they don't feel qualitatively different from GPT-4o's for most users. Sub-300ms streaming voice does feel different. It's the difference between a chatbot and a colleague. Whether that's enough to justify the pricing model, the reliability tradeoffs, and the still-limited vision channel is going to be answered by whichever consumer product ships first with a GPT-Live backbone and gets to a million daily users. Watch for that launch in the next 60 days — the pattern of OpenAI's API releases suggests they've already handed pre-release access to whichever partner is going to use it as a launch platform.
This is the opposite direction AI should be going. Human relationships are the most valuable thing we have, and so, naturally, technology seeks to intermediate and now replace them.I'm not Catholic, but this podcast presents a very interesting argument against talking to AI as if they were huma
What I’m missing from this announcement is the capability to use connectors and tools. I don’t really get it - NONE of the frontier assistants can use tools / connectors while in voice mode - Claude, ChatGPT, Gemini, Grok. It seems so obvious: I want to be able to research stuff, pull up docume
Once this gets video capabilities and is ported to glasses, it'll be a major revolution for blind people (and I say this as a blind person).People have tried "smart <thing> that helps blind people navigate" since the 80s, many, many, many times, and all such projects failed. The
The ad with the grandmas is cute and funny, but from the first 20 seconds you can see that the voice annoyingly interrupts people while they are talking. It's almost as if it tries to reply too fast - faster than a real person would, and the results is that it replies while you're still ta
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I had preview access to this one for a few weeks. It's very good. I had one conversation that lasted a full hour while I was walking the dog, got some good brainstorming done against one of my projects.The best feature is that it can delegate questions out to GPT-5.5 in the background, so you&#