MarkdownModels

Claude Fable 5.1: cache economics for long-running agents

Reviewed: 18/09/2026 · Status: Current public platform information · Audience: Agent platform owners, architects, FinOps and governance

Claude Fable 5.1 is positioned for demanding coding and knowledge work. Its headline token rate is $10 per million input tokens and $50 per million output tokens, but its distinctive TokenOps lever is a $0.25 per million cache-read rate—75% below the preceding Fable 5 cache-read price according to Anthropic.

Published price components

Meter	Price / 1M tokens
Base input	$10.00
5-minute cache write	$12.50
1-hour cache write	$20.00
Cache hit / refresh	$0.25
Output	$50.00

The cache-write premium means caching is not automatically cheaper. It becomes valuable only when a stable prefix is reused enough times before expiry.

Break-even analysis

Let I be the stable input tokens. Compare uncached input with a cache write followed by reads.

For the 5-minute cache:

uncached N calls = N × I × $10/M
cached N calls   = I × $12.50/M + (N-1) × I × $0.25/M

Break-even occurs before the second call:

12.50 + (N-1) × 0.25 < 10N
N > 1.256

So two uses within the valid window already win on input cost. For the 1-hour write:

20 + (N-1) × 0.25 < 10N
N > 2.026

Three uses within the hour are needed.

For a 200K-token stable prefix reused ten times:

Strategy	Input cost
No cache	$20.00
5-minute write + nine reads	$2.95
1-hour write + nine reads	$4.45

This excludes output and any non-cacheable input. The 5-minute option saves 85.3% in this example if the reuse timing is real.

Design an agent for cache locality

Put stable material before dynamic material:

system policy and role;
tool schemas in deterministic order;
repository map or reference corpus;
reusable examples;
conversation or task-specific state;
the latest user instruction.

Avoid timestamps, random IDs, reordered JSON keys and request-specific metadata inside the stable prefix. A single early difference can reduce reuse.

Compaction as a second lever

Anthropic's September 2026 API notes describe on-demand conversation compaction in beta. Compaction replaces older messages with a signed summary block while preserving selected recent turns. Treat it as a controlled state transition:

compact before the context becomes expensive, not only at the hard limit;
preserve unresolved requirements, decisions, constraints and evidence;
compare post-compaction task success against a non-compacted control;
log tokens before and after, compaction frequency and rework rate;
never assume a shorter context is cheaper if it increases tool calls or retries.

Agent budget envelope

Each run needs four independent ceilings:

run budget = input + cache writes + cache reads + output + tool fees

Also cap:

maximum turns;
maximum tool calls by tool type;
maximum output per turn;
maximum wall-clock duration;
maximum failed-action repetitions.

Stop conditions should be deterministic. “Continue until complete” is not a budget policy.

When Fable 5.1 is economically justified

Use it where the accepted outcome benefits from frontier capability and repeated context:

long-running coding agents with stable repository guidance;
research workflows reusing a substantial evidence pack;
professional analysis where error cost exceeds inference cost;
high-value agents whose repeated tool schemas and policy blocks cache well.

Route short, routine extraction or tagging elsewhere unless evaluations prove a material quality advantage.

Operational dashboard

Track these by workflow and prompt version:

Metric	Why it matters
Cache hit ratio	Confirms reuse actually occurs
Cache write amortization	Shows reads per write
Cost per accepted outcome	Includes retries and quality
Output/input cost ratio	Exposes verbose generation
Tool calls per outcome	Detects agent loops
Compaction savings	Measures tokens removed
Post-compaction failure rate	Protects quality

Sources

Validate cache TTLs, eligibility and current prices against the provider documentation before rollout.