Claude Fable 5.1 Pricing: Only the Cache Read Got Cheaper

Claude Fable 5.1 shipped on 1 September 2026 with one price change: the cache read fell from $1.00 to $0.25 per million tokens.

Everything else on the rate card held. That makes this a one-string migration whose payoff rides entirely on a number most teams do not track — their cache-hit ratio.

Four of the five rates did not move

Anthropic publishes five rates per model, not one. A reference guide compiled on 3 September 2026 lists them side by side, in dollars per million tokens:

Rate	Fable 5.1	Fable 5	Opus 5	Sonnet 5
Base input	10	10	5	2
5m cache write	12.50	12.50	6.25	2.50
1h cache write	20	20	10	4
Cache read	0.25	1.00	0.50	0.20
Output	50	50	25	10

Read the two Fable columns against each other and the point is blunt: "Four of the five numbers are identical. The version bump moved exactly one line."

What you need for the swap

The identifiers, per the same reference: model ID claude-fable-5-1, a 1M-token context window, 128K max output, adaptive thinking always on, default effort high, and a reliable knowledge cutoff of June 2026. The migration itself is a string — "base rates, context window, output ceiling and default effort are all unchanged".

One budgeting caveat: thinking cannot be turned off, and "thinking tokens are billed as output — the $50 line, not the $10 line".

Where the cheaper read actually pays

The guide models one 1M-token prefix written once and read back N times, assuming a single 5-minute write with no expiry in between:

Reads	Fable 5.1	Fable 5	No caching
1	$12.75	$13.50	$20.00
10	$15.00	$22.50	$110.00
100	$37.50	$112.50	$1,010.00
500	$137.50	$512.50	$5,010.00

At one read the two versions sit "within 6% of each other". At a hundred reads, "Fable 5.1 costs a third of Fable 5". Nothing about the model changes across those rows — only the cache-hit count does.

Effort raises the bill through volume, not rate

Per-token rates hold at every effort level; what moves is how much the model writes. Artificial Analysis, which the guide says took part in pre-release evaluation, measured Fable 5.1 at max effort using roughly 1.7x the output tokens of Fable 5. Its cost-per-task figure came out at $3.76 against Fable 5's $3.14 — 20% higher — even with a 75% cheaper cache read. If you are cost-sensitive, cap effort before optimizing anything else.

The TTL choice is about idle time, not model choice

Two write durations are sold: $12.50 per million for five minutes, $20 for an hour. Anthropic's prompt-caching documentation, as quoted in the guide, states the cache "is refreshed for no additional cost each time the cached content is used", with the lifetime measured from each request that writes or reads the entry. A continuously busy agent therefore holds a five-minute entry open on reads alone.

The leak is in the gaps. Any pause over five minutes — a slow tool call, a human reading a diff — expires the entry, and "an hour of that costs at worst 12 × $12.50 = $150, against $20 for one 1-hour write".

Is it worth switching?

If your prefix is re-read dozens of times per window, yes, and the change costs a string. If your agent idles, buy the 1-hour TTL first — that decision is worth more than the version bump. And if your volume is large and steady enough that self-hosting is on the table, a separate TCO guide's worked example puts the break-even at 175 million tokens per month, on assumed fixed costs of $1,400 monthly against an $8-per-million API workload; the same guide is clear that APIs stay cheaper "for prototypes, irregular traffic, and low-volume workloads".

Two things to check after the swap

Output-volume alerts calibrated on Fable 5 will fire on that 1.7x figure. And the June 2026 cutoff means prompts written to patch around stale library knowledge "may now be redundant or, worse, actively wrong". One more, for anyone reading launch notes: the 60%-fewer-false-positives claim in the 1 September announcement is a cybersecurity safeguard figure, separate from the roughly 85% biology-fallback reduction Anthropic reported on 7 August — the guide warns against adding them together. Before you bank any of this, instrument usage.cache_read_input_tokens and measure your own hit ratio.