# Resuming after a break costs 10x? On a subscription my breaks cost $12 in 2.5 months. On an API key, about $650 — retained record

> Retained reddit record from reddit:ClaudeAI.

Canonical: https://highsignal.app/data/records/05a5f185a4184664

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeAI)  published 2026-11-17 14:19 ingested 2026-11-18 02:36

[original source ↗](https://www.reddit.com/r/ClaudeAI/comments/1xylmkm/resuming_after_a_break_costs_10x_on_a/)

## retained text



```
https://preview.redd.it/bdy85u53k62i1.png?width=2400&format=png&auto=webp&s=d07f6c8dae74ae6422ab5b056af7b239a9ab9c35

A repo making the rounds this week ([inevolin/claude-code-usage-analysis](https://github.com/inevolin/claude-code-usage-analysis)) found that coming back to a session after 6 minutes costs about 10x. Their numbers check out, but they were on an API key, where the prompt cache only lives 5 minutes. On my Pro sub it lives an hour (99.8% of my cache writes were 1-hour ones), and I'd guess Max and Ultra get the same.

So I ran the same question against my own logs: about 186k model requests since early September, all on a subscription, bucketed by how long the pause before each one was. "Cache rebuilt" means the cache had gone cold and the whole context got written again at full price.

Turns out the cliff is real, and a rebuilt request costs me 14 to 16x a warm one. It just kicks in at the hour mark, not at 5 minutes.

In API dollars (list prices, mostly Opus) my hand sessions came to about $5.7k over those 2.5 months, roughly half my usage. The 17 cold resumes after an hour+ added $12 on top. Chump change, about 70 cents a pop.

What if I'd been on an API key, though? The 5 minute cache is cheaper to write (1.25x input vs 2x), so that alone would've saved me about $600. But every coffee break over 5 minutes would've rebuilt the cache, about 2,000 times all told, for $1,245. Net, the key comes out roughly $650 dearer, 11% more for the privilege of stepping away from the desk

So on a sub you can grab a coffee and not think twice. Lunch is where I'd start fresh with a short note. On a key, anything past 5 minutes is the pricey one.

Queries are here if you want to check my math: [cache-cliff.sql](https://github.com/NamerekT/even-keel/blob/release/v0.3/docs/posts/cache-cliff.sql). They run against my own stats db so the table names won't match yours, but the logic is short.

Anyone on an API key care to run the same split on their logs? And if you're on Max or Ultra, do your cache writes show up as 1-hour too?

Edit: as a few people pointed out, the 5m cache is only the default on a key. You can set 1h yourself (ENABLE\_PROMPT\_CACHING\_1H=1 in Claude Code), so the $650 is really the price of leaving the default alone. Subscriptions get the 1h cache on the main conversation automatically
```



document fetched 2026-11-18 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xylmkm/resuming_after_a_break_costs_10x_on_a
