Claude Sonnet 5 wants to do your errands
Sonnet 5 for agents, Mythos and Fable are free again, Zuckerberg admits agents are late
Hey, sapiens.
Summer is here, and the AI labs did not take a holiday. They are building agents to take ours instead. Anthropic shipped a model made for doing work on its own, the US let two powerful models back out of export jail, and Ford rehired 350 human engineers. Yes, humans. Hold onto that warm feeling.
Anthropic ships Claude Sonnet 5 for agent work
The gistAnthropic released Claude Sonnet 5, a mid-size model built for agent work. It plans, uses browsers and terminals, and can run tasks on its own, according to Anthropic. The pitch is simple: agent power without the big-model bill. Companies running thousands of automated tasks care about price per task, and Sonnet is meant to be the default worker, with bigger models called in for the hard parts. OpenAI and Google pitched similar models weeks earlier, so the fight is now about cost and reliability rather than who got there first. Why it matters: most real AI work happens in the middle tier. When that tier gets much better, automation spreads to jobs that were too expensive to hand over before. What to do: pick one multi-step task you repeat every week and try it with an agent. Note where it gets stuck. That is where your job is safe, for now.
The US lifts export limits on Anthropic's Mythos and Fable
The gistThe US dropped the export licence rule on Anthropic's Mythos and Fable models, with access back on July 1 after a June 12 restriction. Anthropic agreed to security commitments.
Governments now treat top models like strategic goods. Access can switch on and off with policy, which matters if you build on them.
Microsoft bets $2.5B on making companies' AI actually work
The gistMicrosoft launched Frontier Company, a new business with $2.5 billion and 6,000 experts to help clients deploy AI. Amazon, OpenAI and Anthropic have similar ventures.
The bottleneck is no longer the model, it is getting it into real workflows. Consultants are the new GPUs.
Zuckerberg admits AI agents are behind schedule
The gistZuckerberg reportedly told staff that AI agents have not accelerated as executives expected. Meta cut about 8,000 jobs earlier this year and moved 7,000 more into AI groups.
Even the biggest AI spenders are learning that agents are harder than demos. Bad timing for the people already laid off.
Meanwhile in Europe
Proton's privacy chatbot Lumo learns to draw and remember
The gistSwiss company Proton released Lumo 2.0, adding image recognition and generation, a thinking mode and user-controlled memory for projects. A rare AI assistant that sells privacy first.
Enjoy your summer. The agents are working through theirs. Ciao, sapiens.