Anthropic

“Claude 4.7 ships with native long-running agent support. The harness is what they've been quietly testing on Claude Code customers for months. The benchmarks are real but the actual unlock is the cost — same intelligence at roughly half the inference spend.”

“The Claude memory feature isn't what people think. It's not a vector store under the hood — it's a structured event log per user with an LLM-driven retrieval policy. Much more boring, much more reliable. Quiet ship in the next month.”

“Anthropic just leased two more floors at 500 Howard. They're hiring 80 engineers in the next four months, mostly model behavior and infra. Saw the badge order go through the building manager.”

“The new Claude model is noticeably better at refusing poorly-scoped asks. That's a design call, not a capability one. Someone made a choice that safety-at-margin beats user-pleasing benchmarks. I respect it.”