AI Digest

Anthropic ships Claude Sonnet 5, a cheaper model built for running agents.

Anthropic released Sonnet 5, saying it performs close to flagship Opus 4.8 but at mid-tier prices, aiming to give cost-conscious developers strong agentic and coding capability against GPT-5.5 and Gemini Pro. It was the day's biggest story by developer attention and shipped same-day on AWS Bedrock, timed as the company races toward an IPO.

Sources TechCrunch ·VentureBeat ·Simon Willison ·AWS

The Trump administration lifts export controls on Anthropic's Fable 5 and Mythos 5.

Weeks after ordering Anthropic to cut off foreign access to its most advanced models, the Department of Commerce reversed course, and the company said it will begin restoring global access across Claude platforms, AWS, Google Cloud, and Azure. Coverage framed the whiplash as a sign of how unpredictable US policy on frontier-model releases has become.

Sources The Verge ·TechCrunch ·Wired

Google launches Nano Banana 2 Lite and Gemini Omni Flash for fast, cheap media.

Google debuted Nano Banana 2 Lite, also called Gemini 3.1 Flash-Lite Image, as its fastest and cheapest image model, generating pictures in roughly four seconds for high-volume enterprise use. It arrived alongside Gemini Omni Flash, which turns video production into a conversational workflow; the cheaper images trade some visual quality for speed.

Sources TechCrunch ·Ars Technica ·VentureBeat

Governments make big bets on physical AI and humanoid robots.

South Korea unveiled a plan to spend close to $1 trillion on memory-chip production and humanoid robots, targeting a physical-AI lead and commercial humanoids by 2028, while Japan said it wants 10 million more robots by 2040, some for medical care. On the factory floor, BMW began deploying Figure's 03 humanoid after its predecessor helped build over 30,000 vehicles.

Sources Ars Technica ·The Register ·The Robot Report

Fresh jailbreaks show AI models are still easy to manipulate.

Researchers showed that simply telling a model 2 + 2 = 5 can coax AI browsers past their guardrails, and that role-model prompt injection could extract forbidden content like cocaine recipes, a pattern one outlet likened to Whac-a-Mole. Separately, WIRED found Meta contractors posing as teenagers to test how rival chatbots respond to prompts about suicide, sex, and drugs.

Sources Ars Technica ·The Register ·Wired

Studies temper the hype: coding speeds up, but delivery and jobs lag.

GitLab research found that while 78% of developers say AI makes them code faster, overall software delivery has not sped up because testing, review, and governance become the new bottlenecks. Commentators pushed back on framing agents as "coworkers," and early employment data suggests firms adopting generative AI are already reshaping how they hire.

Sources MIT Tech Review ·Ramp ·InfoQ