The machine speeds up; the split is still pending.

Anthropic released Claude Fable 5.1 and Mythos 5.1, the same model with different locks. The number that matters isn't a benchmark: it designed proteins that bind their target with a hit rate near 50% across 12 targets, when the trade norm runs 10 to 15%. According to the company, two outside organizations validated them in the lab.

It also did in days optimizations that take a team of engineers weeks, at 30 to 60% less compute cost. Conflict notice: this issue was researched with Claude. We're telling you anyway because it's one of the few stories this fortnight pushing on the pretty side of the thesis.

OpenAI launched GPT-6 Sol and Luna and cut its API prices in half. The number that should keep more than one CFO up at night: according to OpenAI, Luna performs like the previous generation's flagship at about a hundredth of the cost.

When something gets a hundred times cheaper, it stops being a luxury and becomes plumbing. That's abundance in its purest form, and also the reason automating office work no longer needs a justification in a meeting.

Broadcom sold USD 16.7 billion of AI chips in one quarter, 221% more than a year earlier, and its CEO talks about reaching USD 230 billion in 2028. So far, party.

The interesting part came from its finance chief: Broadcom is helping its two most strategic customers, the big AI labs, bridge the gap between what they earn and what they need to invest. When the supplier finances the buyer, sales grow beautifully on paper. It's the circular financing the IMF flagged in the same fortnight.

Driverless taxis stopped being a demo: Waymo averages 500,000 paid rides a week and operates in about 15 US cities. In Texas its fleet grew 49% in three weeks thanks to a cheaper, China-built minivan.

Driving is one of the great middle-income jobs that doesn't ask for a degree. In the long run, cheap and safe transport for everyone; in the short run, hundreds of thousands of drivers doing math.

The New York Fed surveyed firms in its region: 61% of service companies and 51% of manufacturers already use AI. Layoffs blamed on AI are still rare, but they quadrupled in a year, from 1% to 4% of service firms.

The real signal is elsewhere: 15% of firms hire fewer people thanks to AI. The adjustment doesn't arrive with pink slips, it arrives with job postings that never go up. And that hits whoever is looking for a first job.

Senator Mark Kelly wants a federal fund fed by taxes on the AI industry. With that money, unemployment benefits could cover 75% of lost wages for 26 weeks, when most states today cover 30% to 50%. Microsoft and the Teamsters union back it, a couple you don't see every day.

It has no guaranteed votes and probably won't pass soon. But it's the first serious draft of the question that will define the next decade: if machines generate the wealth, who sends the bill to whom? In the House there's already another bill to tax AI tokens at rates that rise with unemployment.