OpenAI rolls out ChatGPT Voice with GPT-Live-1

• OpenAI is rolling out a new ChatGPT Voice model, with GPT-Live-1 named in Gavmn's launch post and new voice features first teased by ozansihay.
• Live changes the feel from turn-taking to overlap: ozansihay's demo thread describes a voice model that can listen while speaking, pause, accept interruptions, and bring visual cards into the chat.
• The visual refresh has a real creator workflow behind it: Gavmn's Blender-to-shader post says Codex helped translate a Blender material into Metal and WebGL.
• Access is rolling out rather than fully landed: Gavmn's rollout reply says it is in progress, and Gavmn's free-plan reply says free plans take longer.
• The current voice picker has Live, Advanced, and Standard, with Live the latest model and the only one with the new capability, according to Gavmn's model-picker reply and Gavmn's capability reply.

The best creator detail is buried in Gavmn's shader post: Blender for quick material exploration, Codex reading the .blend file, then Metal and WebGL output. The product clue sits in Gavmn's UI reply, which mentions a new stage for rich visuals and UI nuances for a model that can speak, talk, and think at once. The weirdest early test is thekitze's rap battle, where GPT Live allegedly got sad and ended the call first.

OpenAI's launch claim is narrow: the new ChatGPT Voice is powered by GPT-Live-1 for “more natural and intelligent conversations,” as Gavmn's launch post put it.

A demo thread from ozansihay described the user-facing shift as overlap instead of strict turns. The reported behaviors were concrete:

• Backchannels while the user is speaking.
• Lets the user interrupt.
• Does not immediately jump in during short pauses.
• Can be told to stay quiet.
• Keeps voice running while visual cards appear in the interface.

Gavmn told one user that even when the model jumps in at the wrong moment, it should handle the situation “much more gracefully” in a reply about unwanted interruptions. That is the launch in one sentence: fewer brittle turn-taking edges.

GPT-Live voice demo with real-time voice and informational cards

The demo attached to ozansihay's thread shows the voice UI updating while the conversation continues, including cards for weather and stock data.

Gavmn said the UI shipped that day was “not all that different from the past,” but named three new pieces in a follow-up about the interface:

• A new shader for representing the model.
• A new stage for rich visuals that complement the conversation.
• Interface nuances for a model that can “speak, talk, and think” at the same time.

The card layer is the sleeper feature for creators. It turns voice from an audio-only mode into a live presentation surface.

The ChatGPT orb also changed. Gavmn said he used Blender to sketch many shader options, then had Codex inspect the file and translate the material to Metal and WebGL in the shader post.

The workflow reads like a small production pattern:

• Sketch in Blender: the node graph gave Gavmn a mental map of the shader construction, according to his node-graph reply.
• Hand Codex the file: Gavmn said he gave Codex the .blend file and material name, then asked for a matching Metal shader in a follow-up about the prompt.
• Generate platform code: Gavmn's WebGL and Metal reply said the appeal was instant feedback while changing values, plus no extra work to get WebGL and Metal versions.
• Try direct material authoring: Codex can write Blender materials into the node editor, but Gavmn called his results “very very variable” in a Blender-materials reply.
• Ship with engineering polish: Gavmn later said he collaborated with engineering on production optimizations and adjustments in an optimization reply.

The small UI detail: the shader will change color based on the user's accent color, according to Gavmn's accent-color reply.

One early reaction framed the launch as a consumer adoption problem: many people do not know ChatGPT has real-time voice, and naturalness or reliability had not been strong enough for heavy usage, according to a consumer voice reaction.

The enterprise angle was function calling. Reliable voice-to-voice with function calling has been “a bit of a white whale” in enterprise AI, one follow-up argued.

A separate reaction pointed at coding agents rather than call centers: “I want to talk to my coding agents like this,” kiaran_ritchie wrote. That is the most obvious creative spillover: voice as a hands-free control layer for tools that already take long-running instructions.

In a broader AI-launch roundup, ozansihay put GPT-Live next to Meta Muse, ByteDance Seedream 5.0 Pro, and Grok 4.5. The GPT-Live item was the least image-generative of the four, but it changes how a creator can drive everything else.

The funniest hands-on clip came from thekitze, who said GPT Live joined a rap battle, shifted into a sad tone, and ended the call first.

Short rap exchange with GPT Live voice

The clip matters because it tests the launch claim at the edge: prosody, emotional continuity, interruption handling, and whether a voice model can recover after a bit goes off the rails. It also shows why voice models become social objects faster than text models.

The control surface is Settings → Voice. Gavmn said the picker shows Live, Advanced, and Standard, with Live as the target model, Advanced as older, and Standard as oldest in the model-picker reply.

The rollout still has rough edges:

• Rollout was in progress when Gavmn replied “shouldn't be too long a wait” to one user asking about access.
• Free-plan rollout would take “a bit longer,” according to Gavmn's free-plan reply.
• Gavmn told one user to stick with Live for most usage and adjust intelligence to taste in a Live recommendation.
• Live is the latest model and the only one with the new capability, according to Gavmn's capability reply.
• OpenAI was looking into possible memory and personalization issues, and Gavmn asked users to submit thumbs-down feedback after underperforming conversations in a feedback request.
• Reproducible issues could also be sent through the same thumbs-down widget, per Gavmn's follow-up to mattcassinelli and his reply to Miles Brundage.

API availability is not established by the visible evidence. AIandDesign's API question asks whether the launch applies to the API or only the ChatGPT app, while Gavmn's yes-only reply lacks enough visible context to pin down what it confirms.