Prompt Caching ROI: A Developer's Guide to Saving on Anthropic Costs

Anthropic’s prompt cache is simple: cache reads are cheap; full input is not. Cache reads bill at 10% of the input rate; the 5-minute cache write carries a 25% surcharge (verified 2026-07-02):

Model	Input /1M	Cache write (5m) /1M	Cache read /1M
Claude Fable 5	$10.00	$12.50	$1.00
Claude Opus 4.8	$5.00	$6.25	$0.50
Claude Sonnet 5	$2.00	$2.50	$0.20
Claude Haiku 4.5	$1.00	$1.25	$0.10

Bar chart comparing input price vs cache-read price per 1M tokens: Claude Sonnet 5 $2 vs $0.20, Claude Fable 5 $10 vs $1, GPT-5.5 $5 vs $0.50, Gemini 2.5 Pro $1.25 vs $0.125 — cache reads cost roughly 10% of input across providers

The ROI story is also simple: if your workload is stable, caching wins fast — the break-even is a single reuse (one write + one read = 1.25× + 0.10× = 1.35× the input rate, vs 2.0× for two uncached sends).

When ROI is obvious

Long system prompts that rarely change
Repeated retrieval of the same file chunks
Agents that iterate in the same session with minimal prompt drift

When ROI is misleading

One-off prompts or constantly changing instructions
Pipelines that bust cache every turn with unique timestamps or IDs

How to evaluate

Measure cache hit ratio and tokens per session. Plug those into TokenCheat’s caching calculator — compare full input vs cache read pricing for your model tier.