Summary
Claude's caching system charges 1.25× the input rate for writes and 0.1× for reads. Break-even requires ~3 reads per write (ratio ≥ 0.5). A one-month audit of a heavy Max subscription found an amortization ratio of 0.16 — meaning ~40% of the $100/mo subscription went to cache writes that were never read back.
The billing UI shows cache write costs and cache read costs as dollar amounts, but does not surface the ratio. Users have no signal that their workflow is wasteful.
Proposed addition
Add a Cache efficiency metric to the monthly billing summary:
Cache efficiency: 0.16 ⚠ Low — you may be wasting ~40% on unused saves
See: 4 habits that reduce cache waste →
Threshold logic:
- Ratio ≥ 0.5 → no indicator (healthy)
- 0.2–0.5 → yellow + link to tips
- < 0.2 → orange + "You may be wasting up to 40% of your subscription"
Why this matters at scale
Claude has millions of subscribers. If even 20% of heavy users have ratios below 0.2, that represents a significant fraction of inference compute consumed by cache writes that expire unused within 5 minutes. Surfacing efficiency data nudges users toward better workflow habits, reducing wasted compute at scale.
Real-world measurement and analysis: https://github.com/daskuntal75/llm-cost-kit/blob/main/docs/responsible-ai-cost-framework.md
Effort estimate
The raw data (cache_write_input_tokens, cache_read_input_tokens costs) is already in the billing system. Computing and displaying the ratio is a presentation-layer addition.
Reference
Open-source cost tracking kit: https://github.com/daskuntal75/llm-cost-kit
Full issue context: https://github.com/daskuntal75/llm-cost-kit/blob/main/docs/anthropic-issues-submit-ready.md
Summary
Claude's caching system charges 1.25× the input rate for writes and 0.1× for reads. Break-even requires ~3 reads per write (ratio ≥ 0.5). A one-month audit of a heavy Max subscription found an amortization ratio of 0.16 — meaning ~40% of the $100/mo subscription went to cache writes that were never read back.
The billing UI shows cache write costs and cache read costs as dollar amounts, but does not surface the ratio. Users have no signal that their workflow is wasteful.
Proposed addition
Add a
Cache efficiencymetric to the monthly billing summary:Threshold logic:
Why this matters at scale
Claude has millions of subscribers. If even 20% of heavy users have ratios below 0.2, that represents a significant fraction of inference compute consumed by cache writes that expire unused within 5 minutes. Surfacing efficiency data nudges users toward better workflow habits, reducing wasted compute at scale.
Real-world measurement and analysis: https://github.com/daskuntal75/llm-cost-kit/blob/main/docs/responsible-ai-cost-framework.md
Effort estimate
The raw data (cache_write_input_tokens, cache_read_input_tokens costs) is already in the billing system. Computing and displaying the ratio is a presentation-layer addition.
Reference
Open-source cost tracking kit: https://github.com/daskuntal75/llm-cost-kit
Full issue context: https://github.com/daskuntal75/llm-cost-kit/blob/main/docs/anthropic-issues-submit-ready.md