Claude Fable 5.2 doubles Opus 5.5 pricing; a dev.to guide says switch only for hard tasks

What happened

Anthropic shipped Claude Fable 5.2 on November 9, 2026, and the practical question for anyone already paying for Opus 5.5 is whether to move over. A comparison published on dev.to argues that most users should not: Opus 5.5 remains the right default for everyday coding, bug fixing, writing and data analysis, while Fable 5.2 justifies itself only on the hardest jobs — sprawling multi-file changes, long-running autonomous agent tasks, and bugs that Opus repeatedly fails to solve. According to the post, Anthropic's own launch material points the same way, casting Opus as the everyday model and reserving Fable for the most ambitious long-horizon work.

Pricing and specifications

Fable 5.2 costs exactly twice what Opus 5.5 does: $8 per million input tokens and $40 per million output tokens, against Opus at $4 and $20. A fast mode doubles both models again — $16/$80 for Fable, $8/$40 for Opus — which means Opus in fast mode costs the same as Fable at normal speed. If raw speed is the goal, the dev.to author notes, Fable is not the economical route.

Everything else about the pair matches: a 1M-token context window, a 128K-token maximum output, $0.20 cache reads, and a medium default effort setting in Claude Code. In the author's reading, Fable is not a larger model in capacity terms; it is tuned for extended, multi-step work.

Benchmark gaps

Drawing on Anthropic's own launch table — which the post explicitly notes has not been independently verified — Fable 5.2 leads Opus 5.5 on most suites:

• FrontierSWE v2, agentic coding: 69.9% against 62.3%
• Remote Labor Index, long knowledge work: 41.3% against 21.3%
• Terminal-Bench-Science 0.1: 74.6% against 58.7%
• OSWorld 2.1, computer use: 89.3% against 87.2%, a narrow margin

One result goes the other way: on AA-Briefcase Elo, Opus 5.5 scores 1800 to Fable's 1785. The pattern the author draws is that Fable's edge widens as tasks get longer and harder, and nearly vanishes on shorter or computer-use workloads.

What doubling costs in practice

The post prices out a representative job — 500K input tokens, 400K of them served from cache, plus 50K output tokens — at $1.48 on Opus 5.5 and $2.88 on Fable 5.2. Run twenty of those each working day for a month, roughly 440 tasks, and the totals reach $651 versus $1,267. The gap is worth paying, the argument goes, only when Fable completes a difficult task correctly on the first attempt; on routine work it is money burned.

The recommended workflow escalates in steps: start on Opus 5.5 at default effort, raise effort to high and then extra-high if results disappoint, move only that one stubborn task to Fable 5.2, and switch straight back once the hard part is done.

Switching models in Claude Code

Fable 5.2 requires Claude Code 2.1.321 or later, installed with claude update. From there, /model fable switches over and /model opus returns. One picker detail from the post: pressing Enter saves a model as the default for every new session, while the s key changes it for the current session only. Using s for Fable avoids accidentally staying on the expensive model.

Two API changes to watch

The post flags two behaviors that can break existing integrations:

• Thinking is permanently on. Requests carrying thinking: {type: "disabled"} or a budget_tokens value return a 400 error; effort levels are the only lever. Opus 5.5 handles this identically.
• Forced tool choice has been removed. Setting tool_choice to any or to a named tool returns a 400; the workaround is auto combined with a prompt instruction naming the intended tool.

Why it matters

Model choice has become a budget decision, not just a quality one: a 2x token price compounds quickly for teams running high-volume agent workloads. The benchmark pattern suggests the premium only pays off on long autonomous tasks — precisely the kind of work agentic tooling is pushing everyone toward. Meanwhile, the API changes around mandatory thinking and the removal of forced tool choice mean migrations are not purely a pricing question; older integrations may start failing with 400 errors. And since every performance figure so far comes from Anthropic's own table rather than independent evaluation, the dev.to author's advice to validate on your own tasks before committing spend is sound.