Anthropic allegedly cut Claude Code cache TTL from 1 hour to 5 minutes, users say

April 12, 2026
From above black and white of snow on asphalt walkway in city in cold weather
Photo by Plato Terentev on Pexels

What happened

It has been reported that analysis of raw Claude Code session JSONL files spanning Jan 11–Apr 11, 2026 shows Anthropic appears to have silently changed the prompt cache TTL default from 1 hour to 5 minutes sometime in early March. The claim surfaced on a GitHub issue (https://github.com/anthropics/claude-code/issues/46829) and has been discussed on Hacker News. Who flipped the switch? Nobody announced it — at least, not publicly — and that is the core of the outcry.

The technical and cost impact

Prior to the change, Claude Code was writing 1‑hour TTLs for prompt cache entries, which the poster and some developers believe was the intended default; the shorter 5‑minute TTL has allegedly produced a 20–32% increase in cache-creation costs and a measurable spike in quota consumption for subscription users who had never previously hit their limits. That kind of uptick is not trivial: more cache churn means more writes, more billing units, and surprise rate-limit friction for customers. In plain terms — freshness came at the expense of cost and quota predictability.

Why this matters

This is the emotional moment: developers and small teams rely on predictable defaults. Hit a new limit out of the blue and it feels like being nickeled-and-dimed by an invisible hand. The episode speaks to broader industry tensions — speed and freshness versus cost control — and the need for better change communication from platform providers. If the claim holds, a quiet default change created real downstream effects for paying users.

The response and what to do next

The GitHub thread remains the primary public record of the claim; it has been reported that there’s no clear, official explanation in that thread at time of writing. For now, engineers should audit cache TTL settings, monitor quota trends closely, and consider explicit TTLs in client code rather than relying on defaults. Allegedly or not, the lesson is obvious: when defaults move, so do bills and expectations.

Sources: github.com/anthropics, Hacker News