Fable 5.2 is Claude Code's new default Fable: 1M context, $8/$40, and your saved 5.1 pick just moved

Claude Code 2.1.321 shipped on November 9 with a one-line Added entry that does more than it sounds like: Claude Fable 5.2 (claude-fable-5-2) is now the default Fable model, with a 1M-token context window and pricing at $8 per million input tokens, $40 per million output tokens, and $0.20 per million cache reads (changelog). Further down in the same release, under Changed, is the part that actually touches your setup: saved Fable 5.1 model picks were migrated to the fable alias.

If you had picked Fable 5.1 by name, you are now running whatever fable points at, and that is 5.2.

What shipped on November 9

The facts from the official changelog:

• claude-fable-5-2 is the default Fable model, 1M context, $8/$40 per Mtok, $0.20/Mtok cache reads.
• Saved Fable 5.1 picks move to fable: immediately on Pro and Max plans, on Team and Enterprise once your organization allows it. /model claude-fable-5-1 switches back.
• claude remote-control gained a --settings <file> flag that applies one settings file to every session it starts, and a changed file only takes effect after a restart.
• The next day, 2.1.322 added fast mode on Vertex in the global, us and eu regions.

Anthropic says Fable 5.2 is up to 30% faster than Fable 5.1, and the price dropped from $10/$50 to $8/$40, a 20% cut (Claude Code Daily). Coverage of the launch also reports it beats GPT-6 Astra on every benchmark Anthropic published while costing about 25% less (AIToolsRecap). A sibling model, Mythos 5.2, is the same weights without the cyber and biology safeguards and ships only to vetted organizations, so it will not show up in a coding tool's model picker.

Your saved 5.1 pick is now an alias

This is the migration worth understanding, because it will happen again with the next version bump. The changelog line reads: saved Fable 5.1 model picks move to fable. Mechanically, fable is an alias that resolves to the newest Fable, so a pick you made against a specific version now floats.

There is one escape hatch, and it is the same one I recommended when model: haiku stopped meaning one thing: pin exact model ids where runs must repeat. CI jobs, eval harnesses, shared team configs, and any setup where you compare runs across weeks should not ride an alias. /model claude-fable-5-1 pins the old model back. Everyone else can float; the alias gets you the price cut and the speed bump without any action.

What 1M context costs when you fill it

The unit prices are published, so the arithmetic is honest. Fresh input at $8 per million means filling an entire 1M-token window once costs $8 per pass. Output at $40 per million means a 50k-token response costs $2. Cache reads at $0.20 per million mean a cached 800k-token prefix re-reads for $0.16 instead of $6.40 fresh, 40x cheaper per token. That last number is why long sessions with a stable prefix survive at all, and it assumes your prefix actually stays cacheable, which I wrote about here.

Two caveats. The changelog publishes the cache-read price only, not the cache-write price, so do not build a cost model off reads alone. And a bigger window changes behavior: you stop trimming context because you can. Loading everything still costs real money per fresh pass, and a bloated instruction file still degrades attention even when it fits, which is the whole argument behind the 300-line budget.

Fast mode is the quiet budget multiplier

Fable 5.2 ships with a fast mode: 2.5x faster output at twice the cost, toggled in the model picker in Claude Code and billed through extra usage credits, or set with "speed": "fast" in the API (Claude Code Daily). Version 2.1.322 brought fast mode to Vertex as well, and 2.1.321 relabeled the fast-mode rows in /fast, /config and /model to name your current model when it supports it.

A speed toggle you clicked once and forgot is exactly the kind of thing a release like this resurfaces. Twice the cost on the output leg is the difference between a $2 answer and a $4 one.

Limits and resets changed shape too

Anthropic paired the launch with loyalty-based Fable resets: the longer you have been subscribed, the more resets you get, with Ultra and Team Premium at up to 3 and Pro and Team Standard at 1 plus a trial week (Claude Code Daily). Worth knowing before you assume a reset policy from memory.

Also, do not map the 20% list-price cut straight onto your usage limit. Limit accounting is its own thing: when Anthropic cut subagent status-update costs by about 99% per update, the announced limit effect was about 5% of a session's usage, not 99% (I covered that here). Treat $10 to $8 as an API-bill fact, and watch your own limit burn rate for a few days before drawing conclusions. The last time a default flipped underneath everyone, the budget surprises came from usage patterns, not the sticker price (that story).

Four checks before your next session

First, your version, because the migration rides the update:

claude --version
# 2.1.321 or newer means the Fable migration already ran

Second, what is actually selected. Run /model. If it reads fable, you are on 5.2 now, whether you meant to be or not.

Third, your config pins. Aliases hiding in settings files are the ones that surprise you during an eval:

grep -rn '"model"' ~/.claude/settings.json .claude/settings.json 2>/dev/null
# bare aliases to look for: "fable", "opus", "haiku", "sonnet"
# swap them for exact ids where runs must be reproducible

{
  "model": "claude-fable-5-2"
}

If you drive sessions programmatically, claude remote-control --settings <file> is now the clean place to enforce one pinned model across every session it starts. Remember the changelog's edge: if the file changes, no new session starts until you restart remote-control.

Fourth, fast mode. Open the picker, see whether it is on, and decide on purpose. It is 2x cost, not a free speedup.

Which model for which job

Anthropic's own positioning, as summarized by Claude Code Daily: Fable suits ambitious, open-ended technical work, while Opus 5.5 remains the recommended daily driver for most coding tasks. That maps cleanly onto config. Leave the default where it is for everyday sessions, and pin claude-fable-5-2 for the deep-dive agents and subagents where a 1M window and cheap cache reads genuinely pay for themselves.

The pattern to keep is older than this release: defaults and aliases move, and they move quietly, in a Changed list rather than a headline. Exact ids, a 30-second version check, and a peek at the fast-mode toggle are what stand between a routine update and a surprising bill.