Claude Fable 5.2 cut every price 20%, but a task at the API's default effort costs 20% more than on 5.1. On medium it's $3.80, 37% under Opus 5.5 xhigh for nearly the same score.

Fable 5.2 came out on November 9 at $8 input and $40 output per million tokens. Fable 5.1 was $10 and $50. Cache writes, cache reads and batch all dropped by the same 20%, and the 1M window still has no long-context surcharge. Our launch-day post priced it on fixed per-step workloads. This one looks at what a whole agent task costs once the model decides how much to think.

Run it the way the API ships it, on high effort, and an Artificial Analysis task costs $9.08. The same task on 5.1 at high cost $7.54. That's 20% more, not 20% less. Drop to medium and it's $3.80 for a score of 56.4, which is cheaper than 5.1 at any setting and better than 5.1 at any setting too. We'd set medium as the default and make high something a request has to ask for.

Cheaper tokens, more of them

Here is what Artificial Analysis measured for each effort level on both versions. Score, dollars per task and output tokens per task.

Artificial Analysis Intelligence Index, read November 10, 2026, from the Fable 5.2 and Fable 5.1 pages. Accent marks where 5.2 is the cheaper of the two.

At low and medium, 5.2 writes slightly fewer tokens than 5.1 and they cost less each, so the task gets 31-32% cheaper. Anthropic said it would use fewer tokens than 5.1 at medium, and it does.

From high upward it flips. At high, 5.2 writes 97.8K output tokens to 5.1's 66.0K, 48% more. Price 5.1's high run at the new card and it would cost $6.03. The real bill is $9.08, so the extra tokens added about $3.05 and the price cut took off about $1.51. At max, 5.2 costs twice as much per task as 5.1 did.

To be fair to Anthropic, 5.2 at high also scores 14.6 points more than 5.1 at high. You aren't paying more for nothing. You're just not paying less, which is what a 20% price cut sounds like.

What each step up buys

Walk up the effort ladder on 5.2 and price each step by the points it adds:

Our arithmetic on Artificial Analysis' per-effort figures.

Medium to high is a real jump, 5.4 points for a bit under $1 each, and for hard agent work it's probably worth paying. High to xhigh is the step to skip. It costs $5.75 a task for 0.4 of a point, which on a benchmark this size is noise.

The default matters more than any of this. The API defaults to high, while Anthropic says the Claude apps and Claude Code default to medium. So a Claude Code user mostly sees the cheap version, and an API integration that never sets effort pays 2.4x what it would on medium. Check what your SDK call sends.

Before you swap the model ID

Anthropic's migration guide calls the switch "mostly drop-in" and then lists four breaking changes. The one most likely to bite: the prefix check now applies to every account, so a request that edits an earlier turn returns a 400 unless you set drop_block. Fable 5.2 thinking blocks can only be read by 5.2 and Mythos 5.2, so routing a conversation back to 5.1 or Opus 5.5 midway loses them. The computer_20251124 tool errors outside Bedrock.

The old limits carry over. Forced tool_choice, temperature, top_p and top_k all return 400s, thinking can't be turned off, and 30-day retention is required, so zero data retention needs Anthropic's sign-off. The migration guide also warns that a small max_tokens can run out during thinking and return no text at all, which you still pay for.

There's no rush. The deprecations page keeps Fable 5.1 active until at least September 1, 2027. The tokenizer is the same one 5.1 uses, so your input counts won't move. Run your own evals at medium and high before switching, and compare output tokens per request, not just pass rates.

Fable 5.2 has its own model page. Compare it with everything else on the pricing table, or put your own output length into the calculator. For this model the output token count moves the bill more than the price per token does. Our Fable 5.1 post covers how the 0.025x cache read works.