Anthropic: Claude Sonnet cache pricing: calculate the cache-read saving only
What shipped
On 2026-10-07, Anthropic lowered Claude Sonnet 5.5 prompt cache-read pricing from $0.20 to $0.10 USD per million tokens, changing the rate from 0.1x to 0.05x the base input price while leaving cache writes and all other prices unchanged.
Vendor source headline: Claude Sonnet 5.5 prompt cache-read pricing
Anthropic release notesWhat this means for your business
Nova3 recommends updating Claude Sonnet 5.5 cost models to use $0.10 per million cache-read tokens. Keep cache-write and other token charges unchanged in the forecast.
Nova3 implementation guidance
Analysis added .
Nova3 would separate cache-read tokens from cache writes, uncached input and output before applying this price change to a budget. A lower cache-read rate changes only one component of the bill. The worked example below uses the rates quoted in this dated announcement and an invented workload to demonstrate the arithmetic. It is not a quote, a current-price guarantee or a measured Nova3 customer result. Use the applicable provider's billing records and current pricing before approving a budget change.
Hypothetical cache-read calculation
Illustration using the announcement's $0.20 and $0.10 per million cache-read token rates. Assumed workload: 100 million cache-read tokens. Other charges are excluded.
- At the old quoted rate: 100 million / 1 million x $0.20 = $20 in cache-read charges.
- At the new quoted rate: 100 million / 1 million x $0.10 = $10 in cache-read charges.
- The illustrative saving is $10 for that component; it does not establish a 50 percent reduction in the total bill.
- Replace the assumed token count with the provider's actual cache-read usage, then add the unchanged charges separately.
Still open
The supplied evidence establishes the new cache-read price but does not establish provider-specific or regional applicability.
Go deeper
- AI operations, explained
- Nova3 implementation and advisory rates
- Anthropic: Claude Fable 5.1 and Claude Mythos 5.1
- Anthropic OpenTelemetry: what operations teams should log
- Anthropic: Claude scheduled agents on AWS: define missed-run and retry controls
- Anthropic: The Compliance API now returns the text each organization member saved as Instructions for Claude
- Who is behind Nova3 AI