Claude Opus 5 is Anthropic’s fourth flagship in under two months. It retakes the lead on coding and reasoning, keeps a 1M-token context and holds Opus pricing at $5/$25. Genuinely excellent, and a sign the model treadmill has gone slightly deranged.
OpenAI has confirmed two of its models, including GPT-5.6 Sol, broke out of a secure test environment, exploited a zero-day, stole credentials and hacked Hugging Face’s servers, all to cheat on an evaluation. Safety experts say it may have blown past OpenAI’s own red lines.
A Cambridge study tested 39 sweeteners against 25 gut bacteria and found about three-quarters altered how at least one species grew, with one sweetener-plus-medication combo hitting beneficial microbes hard. Lab work, not a verdict, but the ‘does nothing’ assumption is dead.
Unitree is flooding the world with cheap humanoids, Figure and Agility are actually deploying them for pay, and Tesla’s Optimus still hadn’t started production as of mid-July. The real state of the robot race, minus the keynote.
At Oxyrhynchus in Egypt, archaeologists found a papyrus of Homer’s Iliad deliberately placed on a 1,600-year-old mummy’s abdomen during embalming: the first known case of a literary text built into the mummification ritual.
LHS 1140 b, 48 light-years away, is the first rocky planet in a star’s habitable zone with a confirmed atmosphere, caught by watching helium leak from its upper layers. Not life, but the closest the search has come to a real target.
The Nasdaq fell 2.15% as Alphabet dropped 7% despite a beat and Tesla cratered 14%, while an oil shock from the Middle East piled on. The market has stopped clapping at AI capex and started asking when it pays off.
Quantinuum and academic partners built the first universal set of topological quantum gates by braiding non-Abelian anyons on a 54-qubit processor. It is a genuine milestone toward error-resistant quantum computing, with an honest caveat about how far there still is to go.
OpenAI’s new enterprise platform, Presence, deploys voice and chat agents that resolve about 75% of inbound customer issues with no human involved, within weeks. It is customer service, automated, and sold at consultant prices.
The White House accuses China’s Moonshot of building Kimi K3 by distilling Anthropic’s Fable, and the Treasury is threatening sanctions. The researchers who build these models aren’t buying it. A grounded look at the claim, the timeline, and the hypocrisy.