Claude Code cache changes fuel quota complaints

Anthropic reverted the prompt cache TTL for Claude Code from 1 hour back to 5 minutes for some requests, but users reported that usage is actually being consumed faster.

According to Sean Swanson's analysis, Anthropic introduced a 1-hour cache around February 1 and then switched back to a 5-minute cache around March 7. Anthropic stated that this change doesn't necessarily lead to higher costs, and Jarred Sumner explained that if many requests are one-shot calls used only once, the 5-minute cache can actually be cheaper.

Prompt caching avoids reprocessing context that has already been sent, but the cost structure varies depending on the TTL.

• 5-minute cache writes: 25% surcharge over base price
• 1-hour cache writes: 100% surcharge over base price
• Cache reads: about 10% of the base price

However, Swanson said that as a $200/month subscriber, he never hit a quota limit for six months, but the situation changed after March. He argued that this additional consumption is making Claude Code effectively difficult to use.

On top of this, the 1 million token context window is also adding to the burden. Using this large context in Claude Opus 4.6 and Sonnet 4.6 increases cache miss costs, and Claude Code creator Boris Cherny explained that resuming an old session after being away for more than an hour frequently causes a full cache miss.

Cherny said Anthropic is considering lowering the default context window to 400,000 tokens, while letting users opt into 1 million tokens if they want. As large skills, more agents, and background automations have become more common, larger contexts are now routine, and developers believe cache rebuilding or misses may be a major cause of quota exhaustion. At the same time, some users claim that even on the Pro plan ($20/month), they can only send about 2 prompts per 5 hours, raising concerns about both cache bugs and performance degradation.