Anthropic quietly changed how Claude Code bills for its own safety checks this month. Auto mode — the setting that lets Claude Code run autonomously, executing shell commands and taking actions without asking permission on every step — has always run a classifier to screen each action before it fires. Until version 2.1.278, that classifier check was billed as its own separate model request, on top of whatever the actual coding task cost. As of this update, for API and Enterprise users (and on Bedrock, Vertex, Foundry and gateway deployments), the classifier check now runs folded into the session’s own request by default, and isn’t charged separately. A new row in Claude Code’s /status output shows whether a given session is running the check server-side.
Why a billing mechanic is worth flagging
This is a small technical change, but it’s a useful data point on where the cost of agentic coding is actually heading. Longer autonomous sessions — the kind Anthropic’s own 2026 trends report says are becoming the norm — run more of these checks, and each one used to add up. Anthropic stripping that overhead out by default is a direct, if quiet, admission that the meter was running faster than it needed to for teams doing sustained, autonomous, multi-step work. It’s the same direction as September’s Claude Fable 5.1 pricing cuts on heavily agentic workloads: unit costs for autonomous coding sessions keep dropping as the tools mature.
So what
None of this changes the fundamentals of what a project costs to build well — engineering judgement, review discipline and architecture still set the ceiling on quality. But it’s one more small proof point that the economics of AI-assisted delivery keep shifting in the buyer’s favour as the tooling matures, not just the headline model prices. If your last conversation about an AI-assisted build assumed today’s tool costs are fixed, it’s worth another look. See our AI-assisted development approach, or get in touch for a current view on what a build actually costs.