Anthropic took its public offering on the road, Google shipped models built to decide and to read DNA cell by cell, and an airline drew a line on who gets to book a ticket. Here is what caught our eye today.
Anthropic launched its IPO roadshow, with demand already running several times past the shares on offer. The roadshow pitches 825 million shares at 70 to 78 dollars a share for a Nasdaq listing under the ticker ANT, and TSMC’s board approved up to 500 million dollars for Anthropic shares at the IPO price and up to 500 million dollars for OpenAI preferred stock. Why it matters: the two leading private labs are now being priced in public, and the chipmaker at the centre of their supply chain is buying a stake in its own demand.
Google released Gemini Decision 1, a model built only to decide, not to write. Through the new Decisions API it scores up to 1,000 closed ended questions in a single pass, returning probabilities and confidence in 150 to 250 milliseconds, at 0.10 dollars per million input tokens with no output charge, per Google. Why it matters: most agent work is choosing an option, not composing a paragraph, and a cheap, fast decider changes what agents can run all day.
Google DeepMind introduced AlphaGenome 2, which predicts DNA effects in specific cell types. Trained on more than 107 million cell profiles, it moves from tissue averages to individual cell types and states, and is available through its API for non commercial research, not clinical use. Why it matters: disease variants act in specific cells, so reading them cell by cell is how genetic risk becomes precise enough to act on in research.
DeepMind’s AlphaProof Code mathematically verified Signal’s cryptographic core in four weeks. The system wrote more than six million lines of proof code, with roughly 1 percent flagged for human review, per DeepMind, a process that normally takes years. Why it matters: private messaging depends on code everyone trusts but few can fully check, and proof at machine speed raises the bar for verifying critical software.
Runway released Praxis-1 as open weights, a 27 billion parameter world action model for robots. Released under Apache 2.0 with fine tuning software, the video trained model reported 46.2 percent success on RoboLab-120. Why it matters: open weights let robot builders adapt a capable action policy themselves instead of waiting for a closed service, which should speed real world testing.
United Airlines barred third party AI agents from completing flight bookings. Agents may show United flights but cannot complete a purchase on a traveller’s behalf, a policy reported as citing privacy risks, unintended purchases and aviation rules. Why it matters: this is the first clear line from a major airline on agent commerce, and other travel brands will have to decide where their own line sits.





