Claude Code model-switch hooks: gate the model mid-session

Anthropic shipped Claude Code 2.1.251 on August 28 with model-switch hooks — PreModelSwitch and PostModelSwitch. A script can now block, confirm, or annotate a model change while a session is running. That sounds small. It is the first control that treats "which model is answering" as a governed event instead of a dropdown.

What actually happened

From the Claude Code changelog, 2.1.251 adds PreModelSwitch and PostModelSwitch hook events that block, confirm, or annotate a model switch. The same entry gives SessionStart resume hooks two new inputs: session staleness and the estimated re-cache cost of picking that session back up.

Cost visibility landed alongside it. /usage now shows a spend limit bar, with a matching rate_limits.spend_limit field for status line scripts behind a Claude apps gateway. /cost gained a per-session prompt-cache line — hit ratio, misses, tokens re-cached, warm or cold — exposed to status line scripts as a prompt_cache object.

Two other changes matter for anyone running fleets. CLAUDE_CODE_SUBAGENT_MODEL is now a default rather than an override: an agent definition's model: field and an explicit per-spawn model both win over it. And /effort saves a default effort level per model, so switching models no longer drags your effort setting along with it.

Why model governance matters for your business

Model choice is a cost decision your agents currently make unsupervised. A session that starts on a cheap model and quietly escalates is a budget line nobody approved. PreModelSwitch is where you put the rule — deny escalation in CI, require confirmation on a client repo, log every switch to your own audit trail.

It is also a compliance surface. If you told a client their data goes to one model tier, a mid-session switch breaks that promise silently. A hook that annotates every change gives you the record.

The re-cache cost number is the one to instrument. Resuming a stale session is not free — you pay to warm the prompt cache again. Now SessionStart resume hooks get that estimate as an input, so a script can decide: resume this one, start fresh on that one. Wire it into the prompt_cache status line fields and you can actually see where your token spend goes instead of guessing at month end.

Sources: Claude Code changelog.