Summary: Fable 5.1’s discount applies to cache reads, which fell from $1.00 to $0.25 per million tokens, while input, output, and cache-write prices stayed unchanged. The conflicting 25%-cheaper and 18%-more-expensive findings reflect different workloads and effort levels, not different arithmetic. Firecrawl’s 57-run test found more billed output tokens, while Artificial Analysis measured a per-task cost increase from $3.14 to $3.69.
Two numbers about Fable 5.1 are in circulation, and they point at each other. Anthropic puts it at 25% cheaper than Fable 5, and as much as 45% in some cases. Artificial Analysis, measuring the same pair, found it 18% more expensive. Both are correct, and the reason is worth understanding before you switch models.
The rate card moved in one place
Fable 5.1 arrived on September 1, 2026, less than three months after Fable 5 shipped on June 9. It is the same architecture retuned, so the context window, the max output, and the headline input and output prices all carry over untouched: input remains $10 per million tokens, output $50.
The lone difference sits on cache reads, which cost $1.00 per million tokens on Fable 5 and $0.25 on Fable 5.1 — a cut of 75%. Cache writes keep their $20 per million price with no discount at all, and that ceiling is what limits how much of the saving ever reaches your invoice.
Provider-side caching also shapes your routing and privacy decisions, and the economics work the same way there: Zero Data Retention: Enforcing Provider-Side Privacy on AI API Calls covers what gets stored and where.
Why two measurements disagree
Firecrawl’s 57-run test, run on a Claude Team API account and billed at $26.00, showed Fable 5.1 burning more output tokens at every effort level: 1.37x at low, 1.12x at high, 1.30x at max. Under max effort the growth sits in reasoning rather than text — reasoning climbed from 4,205 to 6,725 while visible output dropped from 3,522 to 3,301. So Fable 5.1 shipped 6% less text while the tokens it was billed for rose 30%.
Where Artificial Analysis parts company is the unit of measurement: they price a model by what one completed task costs. Finishing their Intelligence Index took Fable 5.1 140 million output tokens, against 83 million for Fable 5, and their evaluations barely touch caching. Cost per task went from $3.14 to $3.69.
Anthropic’s numbers, by contrast, are drawn from production traffic at default effort, a mix in which cache-heavy agent sessions dominate. The two figures measure different workloads, not different arithmetic.
What the discount looks like on a real bill
Six sandboxed agentic builds are where the cache-read cut surfaced on the invoice. A Fable 5.1 build that re-read 1.69M cached tokens paid $0.42, while a Fable 5 build re-reading 1.34M paid $1.34. What that comes to is 15% to 30% off the bill rather than the 25% to 45% headline, and the difference is that cache writes are charged at full price.
How far the discount stretches depends on how many times a context gets re-read, not on how large that context is. Build a large cache and read it only twice, and nearly none of the saving lands with you.
Effort level is the lever most teams forget
The effort dial runs from low to max, set with output_config: {effort: "low" | "medium" | "high" | "xhigh" | "max"}. Between low, where Artificial Analysis prices Fable 5.1 at $0.77 per task, and max at $3.69, the spread is 4.8x — bought with eight points of Intelligence Index score.
Defaults are not uniform across products: Claude Code starts at High, while Claude Cowork and Claude.ai start at Medium. Anthropic measured its savings figures at default effort and Artificial Analysis measured at max. When costs surprise a team, the usual culprit is an inherited default rather than the model choice.
Reasoning tokens carry the same $50 per million price tag as visible text, and none of them ever show up in the response. Thinking cannot be switched off on Fable 5.1: leave the thinking parameter out entirely, and budget_tokens comes back as a 400 error. Skills and a trimmed global CLAUDE.md shrink visible output, yet neither one changes how much thinking the model does.
How to decide for your workload
Take one recent run and read the usage fields the API returns, comparing cache_read_input_tokens with output_tokens. When cache reads dominate, Fable 5.1 most likely costs you less. When output dominates, those extra reasoning tokens raise the price of every finished task.
The variable that neither party measured is how tightly you specify the task. Give the model a tight spec and Opus 5 won on cost in every one of the Firecrawl runs. Hand it an open brief and Opus 5 averaged $11.83 to Fable 5.1’s $7.00. The room you leave the model to think appears to matter more than which model you pick.
One caveat before you act on any of this: the benchmark figures cited are vendor-run with production safeguards on, and the standard error Anthropic reports is wide enough that any single row inside that band is a wash. Read the direction as the signal, not the decimal.