Three small releases in two days: 2.1.173 through 2.1.175, 16 changes in total. The theme this time isn’t a shiny new feature – it’s something unspectacular that’s still worth gold in larger teams: control over which models are even allowed to run.
Model governance via allowlist
The headline lands in 2.1.175: the admin setting enforceAvailableModels. There was already an availableModels allowlist that let organizations decide which models are permitted – but the Default model could slip past it. That loophole is now closed.
When enforceAvailableModels is on, two things happen. First, the allowlist applies to the Default model as well: if Default would resolve to a model that isn’t permitted, it falls back to the first model on the list. Second, users can no longer widen the managed list through their own user or project settings.
Sounds minor, but it’s exactly the lever compliance teams need. If you work in a regulated industry or want to keep costs in check, you can now cleanly enforce that only approved models run – no escape hatches.
/usage shows where the tokens go
My personal favorite comes from 2.1.174: the VS Code extension now has a usage breakdown in the /usage dialog. Instead of a single total, you see consumption split by cache miss, long context and sub-agents – and crucially, broken down by Skills, Agents, Plugins and MCP, for the past 24 hours or 7 days.
This is the kind of transparency you start craving the moment your setup grows. Which agent is actually eating the most tokens? Which plugin is quietly expensive? Until now you had to guess. Now it’s right there.
The cleanup work
The rest is bug fixes, but a few are worth it. Background sessions were wrongly inheriting ANTHROPIC_* variables from other sessions – fixed. Bedrock GovCloud regions threw 400 errors because the inference prefix was derived as global instead of us-gov – fixed. The /model picker now shows which model Default actually resolves to (Opus, Sonnet – depending on your plan). And exiting right after an interrupted shell command no longer hangs for one or two seconds.
My take: This is an enterprise release, not a toy release. enforceAvailableModels is proof that Anthropic seriously wants Claude Code inside large organizations – this is about governance, not gimmicks. As a solo user, the /usage breakdown is the real win for me: I finally understand where my tokens go. Sounds boring, saves real money.
Sources: Claude Code Changelog, Releasebot: Claude Code, DevelopersIO: Claude Code v2.1.173–v2.1.175