Outperformed Claude in 50K+ web apps.

Enterprise grade security: SOC II, GDPR compliant.

Looks much cooler.

Washington put OpenAI’s newest model on a government leash. OpenAI previewed GPT-5.6 (led by a flagship called Sol), its most capable model yet, and the White House asked OpenAI to limit who can use it, customer by customer, because it crosses the same cyber and bio line that got Anthropic’s Mythos pulled last month. Days later the government cleared Mythos for 100+ approved US institutions, and pressed Meta to submit its models for review too. For the strongest models, Washington now wants a hand on the release valve.

The next Mac costs more, and AI is the reason. Apple raised prices on Macs and iPads, in some cases 15 to 25%, blaming a component crunch it says it has never seen move this fast. AI data centers are buying so much memory that prices have more than doubled since late 2025, with the shortage expected to last into 2027.

A Nobel laureate left Google DeepMind for Anthropic. John Jumper, who shared a Nobel for AlphaFold (the AI that cracked how proteins fold), is leaving DeepMind after nine years for Anthropic. He’s not alone: Gemini co-lead Noam Shazeer recently went to OpenAI, and several more Google researchers have gone to Anthropic.

Anthropic accused Alibaba of the biggest theft of Claude’s outputs it has caught. In a letter to a Senate committee, Anthropic said Alibaba ran nearly 25,000 fake accounts to pull about 28.8 million Claude conversations over 45 days, trying to copy Claude’s coding and agent skills into a rival model. It’s the same “distillation” move Anthropic earlier accused DeepSeek of, just much bigger, and Anthropic is using the disclosure to press Washington for sanctions on the Chinese labs behind it.

Computer use comes to Google’s cheap, fast model. Gemini 3.5 Flash can now do computer use: hand it a screenshot and a goal and it returns the clicks, scrolls, and typing to carry out a task, looping until it’s done across a browser, phone, or desktop. Letting AI drive a screen isn’t new, but it’s normally slow and pricey; the news is that it now runs on a cheap, fast model, which can make more use cases for this type of automation economically feasible.

Sakana’s Fugu is a manager that delegates to other AI models. Japan’s Sakana AI released Fugu, which takes your request and routes it to a team of other companies’ models, deciding which should plan, execute, and check the work, all behind one API. Because it’s really just an orchestration layer, it doubles as insurance against export controls: if one model gets banned or cut off, Fugu swaps in another with no code change. Early testers say the real-world feel doesn’t yet match the hype.

China’s AI video models pulled ahead this week. ByteDance announced Seedance 2.5, which generates a full 30-second 4K clip from a single prompt (most tools still cap out around 10 to 15 seconds), though only in China for now. Alibaba’s rival model, HappyHorse, climbed to No. 2 across the major benchmarks, a spot that opened up partly because OpenAI discontinued Sora in April after it burned roughly $1M a day for almost no revenue.

An AI helped doctors crack 18 childhood disease cases that had stumped everyone. Researchers at Boston Children’s and Harvard ran 376 unsolved pediatric cases (all of them prior dead ends) through OpenAI’s o3; it surfaced leads from the existing research, and specialists confirmed 18 new diagnoses, several of them answers that already sat in the medical literature but had never reached the family’s doctors. The study just published in a New England Journal of Medicine journal, yet it ran on o3, a model from spring 2025: by the time results like these clear medical review, the AI behind them is already a year or two old, which hints at how much more is possible.

Meta read what people were typing from their brain activity alone. In Meta’s Brain2Qwerty v2 experiment, volunteers sat in a brain scanner and typed sentences, and the AI worked out the words purely from their brain signals, without seeing the keyboard or screen, getting 61% of words right on average (78% for the best volunteer) versus about 8% for earlier non-invasive attempts. It’s not quite mind-reading (you still physically type) and it needs a room-sized scanner, so it’s a lab result, not a product, but it’s a big leap for reading the brain without surgery, and Meta open-sourced the code.