OpenAI GPT-Live: the AI that listens while it talks
OpenAI just dropped its first full-duplex voice model, GPT-Live. It listens while it talks. You can cut it off mid-sentence, and it picks right back up.
Hey — it's the Carat team. 🤗 Today we've got OpenAI's real-time voice AI making its debut, ByteDance's image AI that can finally render text, Runway opening doors for developers, and an AI that matches GPT-5.5 at half the price. Let's go.
OpenAI GPT-Live: interrupt it mid-sentence and it still gets you
OpenAI just dropped its first full-duplex voice model, GPT-Live. It listens while it talks. You can cut it off mid-sentence, and it picks right back up.
Until now, AI voice models had to finish talking before they could hear you. GPT-Live? You say "hold on" and it stops. Say "go ahead" and it continues. Response time sits at 300 milliseconds, close to human conversation speed.
It also filters background noise, recognizes emotion, and supports 9 languages. Rolling out to the ChatGPT app now, with API access coming this week.
For creators, this means voiceovers just leveled up. Not just reading scripts anymore, but real-time interactive audio content.
📌 Today's 3 stories
1️⃣ An image AI that doesn't butcher text
ByteDance officially launched Seedream 5.0 Pro. The headline: precise region editing and native text rendering in 14 languages. Posters, product pages, UI mockups, the kind of images where text matters, the garbled letters problem just got way smaller.
It supports single-image conversion and multi-reference image transformation. Already available on fal and Runway, coming to Carat soon.
2️⃣ Runway just opened the door for developers
Runway launched 'Runway Dev,' a media generation API aimed at developers and enterprises. Search, ads, ecommerce, games, you can now generate and edit AI video programmatically.
This is Runway's first step from a creator tool company toward an infrastructure company. API access is rolling out via waitlist.
3️⃣ GPT-5.5-level performance at half the price?
xAI released Grok 4.5. They say it matches GPT-5.5 in performance, at $2 per million input tokens and $6 per million output tokens, roughly half what comparable large models cost. Benchmark results also place it at the top tier for coding.
They also announced a partnership with Cursor, so you can use Grok 4.5 directly in your coding tools.
That's all for today. From an AI that talks and listens at the same time, to a text-perfect image model, a developer video API, and a half-price AI, we covered it all. See you tomorrow with more good stuff. 😄
