Everyone's framing the Claude macOS agent launch as a pricing-power story. The logic goes: if users can offload real computer tasks to Claude, they'll pay more and stay longer, which lifts ARPU without new acquisition. I started there too. But the more I looked at the competitive map, the less the pricing thesis held up.
Here's the problem. Anthropic is building agentic capabilities that route through their cloud at the exact moment on-device inference is pulling the floor out from under that dependency. Apple's M5 runs 30B-parameter models locally in under three seconds. No subscription, no API key. Apple's Foundation Models framework gives developers free on-device inference with a few lines of Swift. The consumer tasks Claude computer use targets (email triage, file management, spreadsheets) are precisely the workloads migrating to local silicon.
Anthropic is pricing into a cloud dependency that the hardware layer is quietly dissolving. I'd watch whether the workloads justifying $20/month still require cloud inference in 12 months. If Apple Intelligence and local models absorb the consumer task base, the retention thesis inverts.