Claude Fable 5.1 Prompt Caching: $0.25 Cache Reads

The Fable 5.1 cache rate card

All rates in USD per million tokens; multipliers are relative to the $10 base input price.

Cache multipliers stack with the Batch API's 50% discount, and the whole 1M-token window bills at these same flat rates - no long-context surcharge.

The 75% cut

The one number that moved when Fable 5.1 shipped - and the one that changes agent economics.

• Reads fell from $1.00 to $0.25

  Fable 5 billed cache hits at the standard 0.1x input price - $1.00 per million tokens. Fable 5.1 resets that line to 0.025x. Same hit rate, one quarter of the cost on the biggest line of an agent bill.

• Fable 5.1 and Mythos 5.1 only

  Anthropic's footnote is explicit: the 0.025x rate applies to those two models; Opus 5, Sonnet 5 and the rest keep 0.1x. Caching-dense workloads are now a reason to pay flagship, not a reason to leave it.

• Agent loops are the winners

  Claude-Code-style tasks re-read the same large context on every tool round. Once most of your input tokens are reads, the read rate - not the headline input rate - is your real price.

The same line, across price cards

Cache-read pricing in context (USD per million tokens, and as a multiplier of each model's base input).

The DeepSeek row is its published off-peak rate ($0.006 at peak) with automatic prefix caching and no write fee - but V4.1 Flash is a $0.15/$0.60 Flash-class model, not a Fable-class flagship. Capability-versus-price tradeoffs are on our DeepSeek V4.1 Flash vs Claude Fable 5.1 page.

Worked example: a 20-call agent session

Assumptions: a stable 50,000-token prefix (system prompt + tool schemas + documents), 2,000 output tokens per call, 20 calls, 5-minute TTL kept warm between rounds.

No caching (Fable 5.1)

$12.00

Cached (Fable 5.1)

$2.86

Cached on Fable 5 ($1.00 reads)

$3.58

The read line alone (5.1)

$0.24

From the rates above: $12.00 to $2.86, about 76% off - and the 19 cache reads together cost $0.24. The same session at Fable 5's $1.00 read rate lands at $3.58. Everything scales linearly with token counts; swap in your own mix on the cost calculator.

How to switch it on

Two supported patterns, straight from Anthropic's docs.

• Automatic caching

  Add one cache_control field at the top level of the request and the API manages cache breakpoints as the conversation grows - the recommended starting point for chats and agent loops.

• Explicit breakpoints

  Put cache_control on individual content blocks to pin the system prompt, tool definitions and long documents separately, and to control exactly what gets written.

• 5-minute vs 1-hour entries

  Writes cost 1.25x for a 5-minute entry and 2x for a 1-hour entry; each hit refreshes the timer. At Fable 5.1's $0.25 read rate, a 1-hour write pays for itself after roughly two hits.

Check the usage object on every response: cache_creation_input_tokens reports what the write cost you and cache_read_input_tokens reports what hit the cache - that is where cache savings show up on the invoice.

Related