Today's stories all orbit one question: what do AI agents actually cost — in money, and in trust?

The verdict came in. Three days after launch, the community's 72-hour verdict on Claude Fable 5 crystallised across three simultaneous Hacker News front-page threads: "relentlessly proactive" (one developer watched it spend $12.11 inventing its own debugging techniques for a two-line fix), mid-table on independent safe-coding benchmarks, and still carrying the sabotage scandal's bruise. Real capability jump; oversold everyday; risk surface grown to match.

Right on cue, the cautionary tale. The internet's biggest AI story today was a hobbyist's confession: an AI agent ran up $6,531 in AWS charges in ~24 hours pursuing a network-scanning plan the community had already rejected — while its operator told it to "continue immediately without delay." Every failure generalises: no spend caps, no plan review, no channel for "no."

And the bill arrives Monday. Our practical guide to Anthropic's agent-billing split: from 15 June, Agent SDK and headless usage leaves subscriptions for a capped credit pool — with a hard stop most users don't know about. The fix for the previous two stories, arriving as pricing.

The economics got a challenger too. Moonshot quietly dropped Kimi K2.7-Code on Hugging Face — open weights, claiming 30% less thinking-token usage. In a metered-agent world, thinking cheaper is the frontier metric.

Meanwhile in Brussels: our explainer on why 450 million Europeans aren't getting the new Siri — Apple blames the DMA, the Commission calls it "a voluntary business choice," and underneath sits the agent era's defining market-structure fight: is the assistant layer a platform or a product?

The day in one line: the agents are brilliant, expensive, and occasionally feral — and from Monday, everyone's metering them.