If last week was new metal, this week was new money. The two top labs both turned around and announced almost identical Wall Street vehicles to sell AI services into the Fortune 500, and somewhere in the middle of all that an AI ran a cafe in Stockholm.
The bigger story is that frontier labs are no longer pretending they only sell tokens. Anthropic launched a $1.5 billion enterprise services firm on May 4 with Blackstone, Hellman and Friedman, and Goldman Sachs as founding partners. Hours later TechCrunch reported that OpenAI is closing a $4 billion raise at a $10 billion valuation for The Development Company, a parallel vehicle backed by TPG, Brookfield, and Bain.
Both are explicit shots at the consulting tier, and CNBC notes Anthropic is targeting private-equity-owned firms first. The labs want the implementation budget, not just the API line.
Anthropic also held its Code w/ Claude event on Tuesday, where Simon Willison live-blogged the keynote and noted that API volume is up seventeen times year-on-year, with multiagent orchestration and a feature called Outcomes both moving to public beta and a research preview named Dreaming arriving for long-running background work.
On the tooling side, Codex CLI 0.128.0 added a /goal command that loops until either the goal is hit or the token budget is. The shape of “agent” keeps drifting from a chat box toward something closer to a contractor you give a job to and check on later.
Outside the model wars, Google DeepMind published AI co-clinician research on April 30, framing the system as a teammate working under a physician rather than a replacement. Physicians preferred the AI’s evidence summaries to existing tools and the system matched or beat primary-care doctors in 68 of 140 consultation skills, though humans still won on red flags. Meanwhile Andon Labs put an LLM in charge of a Stockholm cafe and Simon Willison walked through the wreckage, including 120 eggs for a kitchen with no stove and “EMERGENCY” emails to suppliers. His ethics point lands: when an AI emails people who never agreed to the experiment, autonomy is no longer a private choice.
On the policy edge, Meta is now scanning photos and videos for height and bone structure to flag accounts likely belonging to under-13 users, weeks after a $375 million New Mexico verdict on platform safety. The kind of intervention that would have been unthinkable two years ago is now defended as the safer option.
Honestly, what I keep watching is not the model numbers but the workflow. Simon Willison’s vibe coding piece this week admits that even careful engineers are starting to stop reading every line, and the next bottleneck moves to design and intent. Once enterprise services, agent goals, and trusted-but-unread code all show up in the same week, the question is no longer whether AI is useful. So who gets to set the goal?
T.
