Anthropic Defends Claude Code Cache TTL Reduction as Cost Optimization
A user-led data analysis revealed that Claude Code's cache TTL was reduced from 1 hour to 5 minutes in March 2026, sparking complaints about increased costs for long-running sessions. Anthropic responded that the change was an intentional optimization because 1-hour cache writes are more expensive than 5-minute writes. The company has declined to revert the change, opting instead for a dynamic TTL selection based on request patterns.
Key Points
- Data analysis of 119,866 API calls showed a shift from 1-hour to 5-minute cache TTL starting around March 6, 2026.
- Users reported that the shorter TTL leads to frequent cache expirations in long sessions, resulting in expensive re-uploads and rapid quota exhaustion.
- Anthropic clarified that 1-hour cache writes cost nearly double the rate of 5-minute writes, making 1-hour TTL less efficient for one-shot requests.
- The change was an intentional server-side optimization designed to lower total costs across a diverse mix of request types.
- Anthropic declined to restore a global 1-hour TTL default but fixed a client-side bug in v2.1.90 that affected quota management.
Sentiment
The overall sentiment is strongly negative toward Anthropic. While a few commenters defend specific technical decisions and acknowledge that third-party harnesses can waste compute, the overwhelming majority express frustration with Anthropic's lack of transparency, fear-inducing enforcement policies, and pattern of silent service degradation. The discussion reveals a community that was once enthusiastic about Claude but is now increasingly disillusioned and actively considering alternatives.
In Agreement
- The cache TTL change is technically justified because 1-hour TTL writes are significantly more expensive, and dynamic selection based on expected reuse makes economic sense
- Third-party harnesses often have poor cache management, with specific evidence that some break prompt caching entirely by aggressively trimming context on every turn
- Anthropic handled the OpenClaw situation reasonably by transitioning to extra credits billing with a discount rather than outright banning
- Claude Code is optimized for token efficiency in ways that third-party harnesses cannot match, justifying the subscription pricing model
Opposed
- Anthropic is silently degrading service quality across multiple dimensions including cache TTL, reasoning effort, usage limits, and response length in what appears to be desperate cost-cutting
- The enforcement of third-party harness bans creates a chilling effect where users fear using legitimate features like claude -p for automation
- Users paying $200/month deserve transparency and premium service, not silent downgrades and mysterious account bans with no customer support recourse
- Anthropic should raise prices honestly instead of enshittifying the product — the current approach erodes trust more than a price increase would
- Competition from Codex is catching up while Claude Code becomes more restrictive and less reliable
- Customer support is effectively non-existent, with banned users having no meaningful appeal process