Cost calculator · prices verified Sep 2, 2026

Is Fable 5.1 cheaper for you?

Fable 5.1 cut cache reads by 75%, from $1.00 to $0.25 per million tokens. It also emits about 69% more output tokens to do the same work, and output costs $50 per million. One change saves you money and the other spends it, so which wins depends on the shape of your workload. Describe yours below.

Fable 5 $0.00

Fable 5.1 $0.00

Opus 5 $0.00

Cost to complete the same job, on one million tokens' worth of Fable 5 work.

Cache reads, as a share of your tokens 70%

How to pick this

Context you have already sent and the model re-reads. High for long agent sessions on a big codebase, near zero for one-off prompts. The three shares sum to 100% and fresh input takes whatever is left, so this slider's ceiling drops as you raise the output share - .

Output, as a share of your tokens 10%

What counts as output

What the model writes back, reasoning tokens included. The expensive kind at $50 per million.

How open-ended is the brief? Tight spec

At this setting Opus 5 burns 1.25 tokens for every 1 Fable spends on the same job (1.25×).

Why this dial exists, and where 1.25× and 16× come from

This is not the effort setting. It is how many tokens Opus 5 burns relative to Fable on the same brief, and it exists because cost is volume × rate. Opus 5's rate is known and fixed at half Fable's, so the rate cannot decide the answer on its own - the volume can, and volume is set by how open the brief is. On a tightly specified task we measured Opus 5 at about 1.25× Fable 5.1's tokens - so its half-price rate wins easily. On an open-ended design brief with no clear finish line, our earlier eval measured Opus 5 exploring 16× Fable's tokens (7.95M against 482k), which overwhelmed the rate advantage and made Opus the expensive option. Opus's per-token rate never changes; the volume does. We model that extra exploration as cache reads, since it is mostly the same context being re-read across more turns - scaling the whole token budget instead over-predicted the measured gap by roughly three times.

If 5.1 writes more to do the same job 1.69×

Where 5.1 emits 1,690 output tokens for every 1,000 Fable 5 emits.

Where 1.69× and 1.10× come from

Two reference points. Artificial Analysis measured 1.69× (140M output tokens against 83M) but both runs were at max effort. Our own runs do not yet pin it down: the same three-task suite gave 1.10× on one run and 1.37× on a repeat under identical settings, with individual tasks ranging from 1.02× to 1.62×. Single runs are too noisy to measure a difference this size. A repeated test is underway. Set it to 1.0× to price the cache cut on its own.

The crossover

The whole model is two lines of arithmetic, per million tokens of work:
fable5 = cache×$1.00 + input×$10 + output×$50
fable5.1 = cache×$0.25 + input×$10 + (output×verbosity)×$50
Break-even sits where the cache saving exactly covers the extra output: 0.75×cache = (verbosity−1)×output×50.

The prices behind it

Token type	Fable 5	Fable 5.1	Change
Cache read	$1.00	$0.25	−75%
Input	$10.00	$10.00	no change
Output	$50.00	$50.00	no change
Output tokens to run AA's index	83M	140M	+69%

Per million tokens. Cache-read prices are from Anthropic's announcement and corroborated by Artificial Analysis' blended figures ($7.70 for Fable 5, $7.17 for 5.1, which back-solves to exactly $0.25 at a 7:2:1 mix). Token counts are AA's measurements on Intelligence Index v4.1.1. See the benchmarks page for the full working.

The two claims are measured at different effort levels

This is the part that reconciles them, and neither source spells it out. Anthropic's 25% and 45% savings were, in its own words, "measured at default effort over four weeks of actual usage in August 2026." Artificial Analysis' 69% verbosity jump comes from two max effort configurations: "Claude Fable 5.1 (Adaptive Reasoning, Max Effort)" against "Claude Fable 5 (Adaptive Reasoning, Max Effort)." So the cost saving is a default-effort measurement and the verbosity penalty is a max-effort one. Nobody had published the ratio at the effort levels people actually run, so we tried to measure it: three tasks on both models through claude -p, pinned at --effort low, high and max. The result was not the clean effort story we expected. At suite level the ratio came out 1.37× at low, 1.12× at high and 1.30× at max - non-monotonic - and re-running the identical low-effort suite gave 1.10× where the first attempt gave 1.37×. Individual tasks swung between 1.02× and 1.62×. The honest conclusion from single runs is that run-to-run variance swamps the effect being measured, so we cannot yet confirm or refute AA's 1.69× from our own data. A repeated test (three runs per cell, heavy and light tasks at low and max) is running now and this page will be updated with it. Hit Agent loop, measured verbosity above for that world, and AA blended 7:2:1 for AA's. Method and raw responses are in the repo under test-results/.

What this ignores

Three things, so you can judge how much to trust it. Cache writes and storage are billed separately and are not modelled here. Verbosity is applied only to output tokens, though in a long agent session more output also means a longer history to re-read, which would push a little more of your spend into the cheap cache-read column and favour 5.1 slightly more than shown. And the 1.69× default comes from one benchmark suite, not from your prompts, which is exactly why it is a slider.