Anthropic opened a new app today. It is called Claude Duet. The app turns Claude into a proactive agent. A proactive agent is software that acts before you ask.

Duet sends you a morning brief. It flags emails and calendar items that need you. You can approve one task, or let Duet finish it alone. Anthropic says Duet checks first before any risky step.

Duet connects to over 6,000 apps. Slack, Gmail and Stripe are three examples. It can schedule a meeting or build a sales dashboard. It can also buy supplies through a store's own checkout.

The app runs on Claude Opus 5.5. That is Anthropic's strongest model today. Duet works inside its own cloud computer. It keeps running after you close the app on your phone.

Access starts through a waitlist. Pro and Max subscribers in the United States get first access. Duet costs the same as your existing plan. The agent draws on your normal usage instead.

Anthropic posted the news on X. The post drew 17,391 likes within hours. It pulled 844 replies too. The story climbed to 129 points on Hacker News. Readers left 124 comments on that thread.

Why did this catch fire? Workers already juggle email, chat and calendar all day long. An agent that handles that load alone sounds helpful. Anthropic also named two rivals in its own post, OpenAI's Dots and Meta's Muse. That framing turns Duet into a direct answer to both products.

What does the reach say about the people building AI? Anthropic keeps pushing Claude past chat and into daily errands. The company bets that professionals want help while they step away from their desk. A tool that acts alone needs trust before anything else. Anthropic is asking for that trust one waitlist signup at a time.

Recursive shared its first AI model today. The model is named Flint. Richard Socher, the founder, posted the news on X himself. Recursive built Flint with its own research system. That system is called the Eureka Machine.

Flint writes code for other programs. It runs locally on a laptop, apart from the cloud. On a MacBook Pro it reaches over 125 tokens a second. A token is roughly one word or word piece. That speed beats most comparable coding models.

Flint holds 40 billion parameters in total. Only 2.5 billion of those parameters work on any single task. Recursive calls this design a mixture of experts. The company released Flint under the Apache 2.0 license. Anyone can download and run it for free.

The model leads its size class on coding tests. One test is called SWE-bench Verified. Another is Terminal-Bench. Flint topped both among similar sized models this week.

Socher's post earned 3,279 likes within hours. It drew 323 replies too. The story reached 15 points on Hacker News. On Hugging Face, the model page gathered 150 likes.

Why did this catch fire? A small team built a strong coding model in under four months. Most labs spend far longer, and far more money, on a first release.

What does the reach say about the people building AI? Recursive argues its automated system designed most of Flint. Humans still made a few key choices along the way. That claim asks readers to judge a new kind of lab. In this lab, software helps build the next piece of software. Coders now test Flint on their own laptops. They want to see if that claim holds up. Early reviewers call the results strong for such a young company.

OpenAI shared a new science result today. Thousands of AI agents sorted data from millions of mouse brain cells. The work used a model called GPT-Rosalind. That model is built for research in biology.

Scientists at Scripps Research first switched off genes one at a time in mice. They recorded how each change affected different brain cells. A human researcher doing that work alone would need months. Some parts might take a full year.

GPT-Rosalind agents finished the sorting in 32 minutes. Each agent studied one gene change on its own. It searched past studies and proposed a reason for what it saw. One finding centered on a gene named DCTN1. That gene links to Perry syndrome, a rare brain disease. Disrupting it raised cholesterol related activity in mouse neurons.

Human scientists at the lab reviewed the agents' reports. They preferred the agents' work over a single researcher's report most of the time. OpenAI posted the finding on X. The post earned 4,594 likes so far. It drew 303 replies too. The story reached 106 points on Hacker News. Readers left 53 comments on that thread.

Why did this catch fire? Readers want proof that AI can speed up slow, careful science. Thousands of agents did in half an hour what took researchers a year before. Brain disease research moves slowly, and families watching it want faster answers.

What does the reach say about the people building AI? OpenAI keeps pointing its models at problems past chat and code. The company partnered with a drug research firm and a hospital for this one. That partnership signals where OpenAI wants credit next, inside labs and hospitals. The result still needs a full peer review before any doctor trusts it.