Anthropic, Introducing Claude Fable 5.1: cache reads cut 75%

Why this is on the rung at all — the boundary, stated

On 2026-09-13 Connor rejected two Anthropic and Google pricing-policy items for this rung (Sonnet 5's cancelled price rise; Gemini Flash's scheduled 2× step-up) as drift back toward the retired Token Price Index and rug-pull log: a vendor moving a price, with no mechanism behind it, is not what this rung collects any more (docs/context/kb-held.md).

This source is close to that line, and it is admitted only because the product being cut is itself the mechanism. A cache read is what a provider charges when a request re-uses context it has already processed and stored: the saved working state of the model (the key-value, or KV, cache) is loaded instead of re-computed. Cutting that one line, and only that line, is a statement about what serving stored context costs, which is the question the "KV cache compression cuts inference cost 50%" thesis asks. If a future Anthropic move is a plain list-price change on input or output, it falls on the rejected side of the line, not this one.

It is still a bare price cut in one respect: Anthropic gives no engineering reason for it.

What the page says (read 2026-09-24)

From the "Availability and pricing" section, verbatim:

"Claude Fable 5.1 is priced at $10 per million input tokens and $50 per million output tokens. Cache reads now cost $0.25 per million tokens, 75% less than Fable 5, which reduces the cost of typical workloads by an estimated 25% and highly agentic workloads by up to approximately 45%."

And: "For workloads that need to run in the US, US-only inference is available at 1.1x pricing for input and output tokens."

The page's announcement list dates the release: "Introducing Claude Fable 5.1 — Sep 1, 2026."

The numbers, checked against Anthropic's pricing page (2026-09-24)

Per million tokens	Fable 5	Fable 5.1	Change
Input	$10	$10	none
Output	$50	$50	none
Cache write	$12.50	$12.50	none
Cache read	$1.00	$0.25	−75%
Cache read as a share of the input price	1/10	1/40	—

For comparison, on the same pricing page: Opus 5 reads cache at $0.50 against $5 input (1/10), Opus 5.5 at $0.20 against $4 (1/20), Sonnet 5 at $0.20 against $2 (1/10). Fable 5.1's cache-read discount is the deepest on Anthropic's list.

What it does to a real bill — worked arithmetic

Anthropic does not say what "typical" or "highly agentic" means. A worked example, our arithmetic on the list prices, with an assumed mix, shows why the saving grows with how much context an agent re-reads:

An agent step sends 1,000,000 input tokens, of which 900,000 are context it sent before (cache reads) and 100,000 are new, and gets 20,000 tokens back.

	Fable 5	Fable 5.1
New input, 0.1M × $10	$1.00	$1.00
Cache reads, 0.9M × $1.00 / $0.25	$0.90	$0.225
Output, 0.02M × $50	$1.00	$1.00
Total	$2.90	$2.225 (−23%)

The more of each request is re-read context — which is what long-running agents do, turn after turn — the closer the saving gets to the cache-read cut itself. That is the mechanism behind Anthropic's "up to approximately 45%", even though Anthropic does not show its sums.

Through-line: how this moves a number on the path from token price to task price

The headline token price did not move at all. The price of a task did, because an agent's task is mostly re-reading its own growing context, and that is the line Anthropic cut. This is the clearest case on the rung of a price sheet where the per-token headline says nothing about the per-task cost: two Fable versions with identical $10/$50 rates can differ by a quarter or more on the same agent job.

A customer quote on the same page points the same way (vendor-selected, so weak evidence): Cognition says Fable 5.1 "matched or edged out Fable 5 in our testing at a lower cost per task, and with the new cache read pricing a Fable-class model is finally economical for the workloads we'd kept on Opus."

Limitations

• Vendor source, no cause given. Anthropic does not say whether the cut follows a cheaper way of storing or loading cached context, or is a pricing choice.
• Savings are Anthropic's own estimates, with no workload mix; the worked example above is illustrative only.
• The page can change. It is a product page, not a dated post. A Wayback snapshot from 2026-09-24 05:50 UTC exists (see archived_copy above); this session could not open it or trigger a fresh save, so the quote above, the fetch date and the page hash are the record.
• List prices. Batch (50% off, per the pricing page), US-only inference (1.1×) and negotiated rates are not reflected.

---

Source: Claude Fable, Anthropic ("Introducing Claude Fable 5.1", 2026-09-01), with prices confirmed on Anthropic pricing. Both read 2026-09-24.