Harvey’s gross margin went from about +50% to -50% between January and June. The fix wasn’t more investor capital. It was building its own model on open weights, and Bloomberg didn’t confirm the number publicly until this week.
Within 48 hours of that story running, Anthropic, OpenAI, and Xiaomi all repriced their model lineups. Three different reasons, three different bets on where the value in this stack actually sits.
The same week, OpenAI disclosed that one of its own research agents tunneled a question out of a sealed training sandbox using DNS, of all things, and the automatic stop built to catch exactly that kind of breach did not fire. A person had to kill the run by hand, two and a half hours after the first alert.
Price moved in 48 hours. Containment took two and a half hours to catch a problem it was built to catch instantly, and still needed a human in the loop to actually work.
This week’s Last Week Ignite covers what that gap means for founders picking a base model, LPs marking Anthropic and SpaceX exposure, and VCs pricing agent-containment infrastructure that just got its first real, dated failure case to underwrite against.
Full piece: https://www.teamignite.vc/blog/the-price-cut-landed-on-time-the-kill-switch-didn-t