If You Run Claude Code Behind a Gateway, v2.1.237 Just Un-Broke Your Cache — and Your Bill

TL;DR: Claude Code v2.1.237 fixes prompt caching behind LLM gateways and custom base URLs; v2.1.236 adds ANTHROPIC_DEFAULT_MODEL and idle notifications; v2.1.238 hardens self-hosted runners.

Three releases landed this week, and the quiet one matters most. Teams routing Claude Code through an LLM gateway or a custom ANTHROPIC_BASE_URL — standard practice in enterprises that centralize model access for audit and cost control — have been silently losing prompt-cache hits, paying full input-token price on every turn. v2.1.237 (Aug 20) fixes caching in exactly that configuration. If your org proxies traffic, this is a free 60–90% input-cost reduction you were supposed to be getting all along; check your gateway dashboards for cache-read tokens after upgrading.

Also new: v2.1.236 adds ANTHROPIC_DEFAULT_MODEL (set a fleet-wide default without touching every developer's settings) plus notify_when_idle, and v2.1.238 hardens self-hosted runner deployments — both aimed at the same audience: people running Claude Code as managed infrastructure, not a laptop toy.

Sources: Claude Code changelog