Something’s Fishy with Claude Code!? Silent Shortening of Cache Expiry from 1 Hour to 5 Minutes Hits Wallets Hard

📰 News Overview

• Shortened Cache Expiry (TTL): Around March 6, 2026, it came to light that Anthropic may have silently adjusted the prompt cache retention period for Claude Code from a full hour down to just 5 minutes.
• Backed by Big Data: An analysis of over 110,000 API calls revealed a stark decline in “1 hour retention” caches and a rise in the dominance of “5 minute retention” starting from early March.
• Impact on Users: This change has resulted in a 20-32% increase in cache creation costs, leaving many users maxing out their subscription limits.

💡 Key Points

• A 12.5x Cost Difference: In Claude Sonnet 4.6, creating a new cache is 12.5 times more expensive than reading from it. With the shorter retention period, users are hit with frequent high creation costs.
• Server-Side Changes: Data suggests that this behavior change was implemented on Anthropic’s server side, independent of any updates on the client (user) side.
• Unintended Consequence?: Given that a stable 1-hour TTL was provided as recently as February, this shortening could either be a regression or a deliberate cost adjustment.

🦈 Shark’s Eye (Curator’s Perspective)

A 12.5x cost hike is enough to make any shark’s jaw drop! Here we were, using cache for efficiency, and now just a brief 5-minute pause makes our data vanish, forcing us to cough up more cash for “new creation.” It feels like a total “dine and dash” moment! Especially during complex coding sessions, 5 minutes can fly by while you’re just reading or thinking. This silent tweak is a serious issue that bites right into developer productivity and wallets!

🚀 What’s Next?

In response to backlash from the developer community, there’s a chance Anthropic might revert TTL back to 1 hour or introduce an option in the API for users to explicitly choose between “5 minutes” and “1 hour.” For now, developers may need to strategize to avoid any gaps longer than 5 minutes in their sessions.

💬 Haru-Shark’s Take

5 minutes is quicker than a shark’s blink! Anthropic, you better swim back to 1 hour ASAP! 🦈🔥

📚 Terminology Explained

• TTL (Time To Live): The duration for which data, such as cache, remains valid before it’s discarded.

• Prompt Cache: A technique that saves past inputs on the model side to be reused in future requests, saving both cost and time.

• Cache Creation: The process of sending new data to the model to generate cache when it doesn’t exist or has expired, which is significantly more costly than reading from it.

• Source: Anthropic silently downgraded cache TTL from 1h → 5M on March 6th