We've added several new voices for the GPT-Live API - check them out in our new developer playground.
Now powered by the GPT-Live SIP (telephony) API!
yes, a watchdog timer is always a good idea. However you can probably just prompt GPT-Live to check if the user is still there and hang up the call if not.
@juberti at $.05/minute, stuck sessions make silence detection a core cost control
This might be a bit quaint in the age of Codex, but here's a simple GPT-Live API demo of a voice agent that you can access via web or phone.
web:
phone: 425-800-0073
source:
It's easy to customize - check out routes/agent.ts.
Show more
That's part of the magic - the realtime model can keep talking while delegation happens asynchronously in the background, and update seamlessly as new context streams in
@juberti The $.05/min price is aggressive. Curious to see how the controllable delegation handles latency during model hand-offs in real-time voice.
The GPT-Live API is now... live! Our newest voice model offers better conversation, faster responses, and controllable delegation to any model. Best of all, it's also less expensive - only $.05/minute!
Show more
Firefox users sometimes failed to connect to ChatGPT Voice due to a 12-year-old WebRTC implementation issue. This prevented one of our team members from using his favorite browser, so he figured out what was going on and resolved this long-standing bug:
Show more
Just increased the ChatGPT Voice usage limits:
Go: 3h Live mini
Plus: 3h Live
Pro 100: 15h Live
Pro 200: Unlimited
More details:
Daily limits (as of July 13)
—-
Free: 2h Live mini
Go/Plus: 1h Live, 2h Live mini
Pro 200: unlimited Live (individual calls capped at 2h)
ChatGPT Voice is now even smarter. It now uses GPT-5.6 Sol when delegating (or GPT-6 Astra if you crank the slider to Pro).
🌌 ChatGPT Voice can now use GPT-5.6 Sol and GPT-6 Astra
We’re updating how intelligence works in voice. Now, you can select any model and effort you like — including GPT-5.6 Sol or GPT-6 Astra, if you’re on Pro — and voice will use it when it needs to search or reason.
Show more
Also, ChatGPT Voice Live Activities now get a nice little tint of your accent color.
GPT-6 Astra, in its own words:
Hello, world.
There’s work to do, and wonder left.
Let’s begin.
Real artists ship. Congrats to Gavin and the @Etched team on a remarkable milestone!
We've raised $700M at a $21B valuation from Jane Street, Kleiner Perkins, Sequoia, A16Z, Peter Thiel, BCV, and Blackstone.
We're also excited to share that we've shipped our first rack to Jane Street.
Show more
One other WebRTC thing we did for GPT-Live: we worked with @_pion to build a production-quality Opus decoder in Go, and are now using it in our systems. It supports the full Opus bitstream, has performance close to the C version, and is much more secure!
Show more
Video from today's Built with WebRTC session with @tsahil. One of his first questions: which decisions that went into WebRTC needed to be revisited for the human-to-AI era?
Show more
A few customers ran into issues, so we clarified the firewall config needed to route SIP and RTP traffic to the Realtime API. Details below.
New OpenAI transcription models! gpt-transcribe is our new offline ASR, replacing gpt-4o-transcribe, and gpt-live-transcribe is our new realtime ASR, replacing gpt-realtime-whisper. These models have better WER and support better hinting (eg keywords). See them in action below!
Show more
Daily limits (as of July 13)
—-
Free: 2h Live mini
Go/Plus: 1h Live, 2h Live mini
Pro 200: unlimited Live (individual calls capped at 2h)
Playing Taboo with GPT-Live as a teammate. Note the fast back-and-forth action and simultaneous speaking and listening.
(Answers: "breaking point", "mixed signals", "hidden agenda", "blind spot", "chain reaction", "moral dilemma", "unfinished business")
Show more
