Fable 5.2 landed on November 9 with better coding, reasoning and vision, roughly 30% faster generation, and a price cut from $10/$50 to $8/$40 per million input and output tokens. If all you did was switch the model and carry on, you left most of the release on the table. Three commands shipped alongside it and they solve problems you actually feel.
/model: read the quota rule before you switch
The command is trivial, and it needs Claude Code 2.1.321 or later:
/model claude-fable-5-2
Anthropic's own advice is blunt: Fable 5.2 suits large, open-ended work, while everyday coding should still start on Opus 5.5. A faster model is not automatically a cheaper one — changing a line of CSS does not benefit from 30% more throughput, but you pay the new rate for it.
One date matters for the quota. For Pro and Team Standard accounts, before November 16, Fable counted inside the weekly allowance up to 50% of that allowance; anything beyond it drew on usage credits if you had them enabled. Check the official note in your own account for what happens after that date rather than budgeting from a blog post.
/storage: when the disk fills, start with old worktrees
Old sessions leave worktrees behind, and a few months of them is several gigabytes. You can now look and delete from inside a session:
/storage
Settings shows the same breakdown per session and worktree. Be clear about the boundary: what goes is the unused worktree; transcripts and branches stay. Uncommitted changes go with it. Commit first — there is no undo.
/usage: see what ate the quota, then let it fix what it can
/usage shows where the allowance went. The new version does not stop at numbers — press f and Claude Code sends a prompt of its own to repair the parts it can, usually oversized context, files re-read on every turn, or sessions that ran far longer than they needed to.

Three smaller changes worth knowing
ultrathink is back with a different shape: it applies xhigh reasoning to a single turn instead of changing the session setting. Use it on the turn that is stuck; leaving it on permanently just burns tokens.
Long conversations move faster. Requests used to resend the whole transcript; most now send only the new messages. When you switch models mid-conversation, the current model summarises first if the transcript is long. And on desktop, queued messages survive a quit, an update or a crash.
Cloud sessions can configure their own environment. On Pro, Max and Ultra, if a session is missing a dependency, a network permission or an API key, Claude lists what needs to change and waits for approval; you enter the key in a form it cannot see. The environment is saved and reused the next time you start a session.
Run them in this order
/storagefirst. Look at the disk, delete the worktrees you are sure about (after committing)./usagesecond. See where the allowance went, pressf, let it try.- Then take a genuinely large, open-ended task, run
/model claude-fable-5-2once, and compare speed and rework against Opus 5.5.
Judge step three on cost per finished task, not cost per token. A fast model that needs fewer retries can be cheaper at a higher rate; a slow model that gets it right first time is genuinely saving the difference.
These are not money-saving tricks
/storage and /usage recover disk and wasted allowance, not model spend. Two things still decide the bill: how long the context gets, and how many times a task has to be redone. The new commands let you see both. Whether they change depends on how you work.