Managing Claude Code Organization Costs: Cache TTL Optimization and Amazon Bedrock Attribution Strategies

Prompt Cache TTL and Invalidation Conditions

Claude Code costs are determined by the hit rate of prompt caching. Cache reads are billed at approximately 10% of the standard input price, while cache writes are a one-time investment costing 1.25 to 2 times the standard rate depending on the Time To Live (TTL). Subscription plans apply a default 1-hour TTL, but switching to usage credits may automatically change this to a 5-minute TTL, so you must adjust the duration using the ENABLE_PROMPT_CACHING_1H=1 environment variable.

The cache is valid only when the request prefix matches exactly. Actions such as switching models via /model, changing /effort, disconnecting MCP servers, adding tool deny rules, running /compact, or upgrading Claude Code change the prefix and invalidate the cache. In particular, resuming a long session immediately after an upgrade causes the entire conversation history to be reprocessed, leading to a sharp increase in costs; therefore, it is recommended to clean up sessions before upgrading.

Organizational Cost Control and Attribution

The points of cost management for an organization vary depending on the access method. Teams/Enterprise plans control spending through spend reports in organization analytics and usage credit limits. Claude Console (API) limits total spend and traffic using per-workspace spend limits and rate limits.

When routing through cloud providers such as Amazon Bedrock, data does not appear in the Anthropic dashboard. To attribute usage per user, you must utilize metric export via OpenTelemetry, a self-hosted Claude apps gateway, or an LLM gateway like LiteLLM. However, note that if cache_control marking is not properly passed through the gateway, cache efficiency may decrease, so caution is required.