Issue #5 · July 3, 2026

Text 122 (+0) · Audio 111 (+2) · Video 122 (+2) · Image 100 (+0) · Agents 117 (+0) · Robots 100 (+0) · Science 100 (+0) · Other 106 (+2)

What became real

Audio+2Jul 1

Hugging Face and Cerebras enable real-time voice AI with Gemma 4

A Hugging Face blog describes running Gemma 4 on Cerebras hardware to power low-latency, real-time voice AI. Faster inference makes conversational voice agents more responsive for support and phone-style interactions.

What still breaks: The post's exact public-access terms and whether it is more than a reference integration are unclear from the payload.

SOURCES · huggingface.co

Other+2Jun 29

Claude models reach general availability on NVIDIA GB300 in Microsoft Azure

Anthropic's Claude models in Microsoft Foundry, running on NVIDIA GB300 Blackwell Ultra GPUs, are now generally available to Azure enterprises. Azure-native businesses can build agentic Claude workloads on newer, higher-throughput hardware without leaving their cloud.

What still breaks: The post is vendor-authored and does not give independent latency, cost, or throughput figures.

SOURCES · blogs.nvidia.com

Video+2Jun 30

Google puts faster, cheaper image model (Nano Banana 2 Lite) and Gemini Omni Flash in the API

Google made Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) and Gemini Omni Flash available to build with; the image model is billed as the fastest and cheapest Gemini image option. Businesses can generate images at higher speed and lower cost for high-volume tasks like product mockups and marketing variations.

What still breaks: It still makes text/spelling errors in generated images, and this is an image, not video, capability.

SOURCES · deepmind.google · simonwillison.net

The Threshold