A recent discovery by users has brought to light a significant, unannounced change in Anthropic's service: a silent downgrade of cache TTL (Time-To-Live) from one hour to just five minutes. This adjustment, reportedly occurring around early March 2026, could be impacting developer quotas and inflating operational costs for anyone integrating with Anthropic's APIs.
The Silent Shift
The issue, documented in a GitHub repository, points to a regression where the cache TTL was dramatically shortened. For developers and applications that rely on client-side or proxy caching to manage API call frequency and optimize resource usage, such a change is more than just a minor tweak; it's a fundamental shift in how data freshness and API interaction are managed.
Previously, responses might have been considered valid for up to an hour, allowing applications to serve cached data and reduce the number of direct API calls. With the TTL reduced to five minutes, applications are now forced to re-fetch data much more frequently. This directly translates to an increased volume of API requests.
Understanding the Impact: Quotas and Costs
The most immediate and tangible consequences for developers and businesses are 'quota and cost inflation'. Here's a breakdown of why this matters:
- Increased API Call Volume: Applications designed to cache responses for an hour will now hit the API 12 times more often for the same data point within that hour, assuming a continuous need for fresh data.
- Faster Quota Consumption: Higher call volumes mean developers will reach their API rate limits and daily/monthly quotas much quicker than before. This could lead to unexpected service interruptions or require costly quota upgrades.
- Elevated Billing: Most API services, including AI models, bill based on usage (e.g., per token, per request). A 12x increase in requests for the same logical operation will directly lead to significantly higher operational costs, potentially without any corresponding increase in application functionality or user value.
- Performance Implications: While not directly mentioned, increased API calls could also introduce more network latency and processing overhead, potentially impacting application performance if backend systems aren't scaled to handle the sudden surge.
What Developers Should Do
For anyone building on or integrating with Anthropic's services, especially those sensitive to API usage and costs, here are some actionable steps:
- Review Usage Metrics: Immediately check your Anthropic API usage logs and dashboards from early March 2026 onwards. Look for any uncharacteristic spikes in request volume that don't correlate with increased user activity or new feature deployments.
- Examine Billing Statements: Scrutinize your recent billing statements for an unexpected increase in charges related to Anthropic API usage. This could be a clear indicator of the TTL change's impact.
- Adjust Caching Strategies: If your application implements client-side or proxy caching for Anthropic responses, you may need to re-evaluate your caching logic. Consider:
- Shorter Cache Durations: Align your internal cache TTLs with the new 5-minute external TTL.
- Proactive Refresh: Implement strategies to proactively refresh cached data just before expiration, managing the load more gracefully.
- Conditional Requests: If supported by the API, use conditional requests (e.g.,
If-None-Match,If-Modified-Since) to avoid re-downloading unchanged data, though this still counts as an API call.
- Monitor Official Channels: Keep an eye on the original GitHub issue or official Anthropic communication channels for any statements, explanations, or potential reversions of this change.
This incident underscores the importance of robust monitoring for third-party API dependencies and the potential financial and operational impact of unannounced infrastructure changes. While the specific API or service affected wasn't detailed, the principle applies broadly to any interaction with external services where caching plays a role.
This article is based on information from a GitHub issue and does not constitute an official statement from Anthropic.