Claude Code marks its messages with caching points. As a result, almost every request consists of a long cached prefix and a new, truncated final part. Based on practical experience with Opus 5.5, reading from the cache appears to cost $0.20 per million tokens, accounting for about 92–95% of the tokens charged for prompts. Meanwhile, a five-minute cache entry costs $5, which is 25 times more expensive than reading from the cache. When an agent needs to run a build or a longer training cycle, they face the problem of the cache expiring. The next request then has to write the prefix to the cache again. We saw this 33 times in a cycle of 89-task runs without compression. As a result, 927,000 tokens were rewritten, which is approximately one-third of the volume billed in full.
Until now, we’ve deliberately left the Claude Code text uncompressed because overwriting any byte within a cached prefix would very likely cause it to become invalid. But here’s what we came up with instead:
– Compress only the uncached part. On the first transmission, only the new part is compressed. Everything already cached stays byte-for-byte unchanged. Fragments shorter than 4,000 characters are sent in full, since compressing each fragment would be much more expensive than sending the entire dialogue as is.
– Keep the cache active while the tools are running; don’t let it expire. While the agent waits for a tool response (model building/training), Kompressor resends the last request every 4.5 minutes with max_tokens: 1, so Anthropic re-reads the cached dialogue, and the five-minute timer restarts. Pings stop on the agent’s next request, after ~50 min, on the first failed attempt, or if a ping finds nothing in the cache.
– Switch to the one-hour cache only when necessary. After the first pause longer than 5 min, Kompressor requests a one-hour cache for that conversation.
We ran Claude Code on the Opus 5.5 model through the Terminal-Bench 2.1 benchmark (89 tasks) and calculated the cost based on Anthropic's actual billing, including pings. The cost per run decreased from $35.35 to $28.30 (-20%). The percentage of tasks solved by the agent without compression was 89.9%, and 90.6% when using Kompressor. The number of tokens lost due to cache expiration decreased from 927,000 to 13,000.
But there’s a caveat: the agent never solves the same task twice in exactly the same way, and the entire difference came down to just 4 out of the 89 tasks. We expect these savings to persist from run to run: about 11% from keeping the cache up to date, and a few more percentage points from compression. This effect is most noticeable on tasks where the agent waits a long time – more than five minutes – for a command to execute: model training, compilation, and long simulations. In those cases, costs decreased by 34%. To try Kompressor, see the GitHub repo below. Setup is in the README. The default configuration already includes compression of new text only and a switch to an hourly cache on long pauses; “keep-warm” is off (the only request that the proxy sends on its own). To enable, add to your config:
[proxy.providers.anthropic]
cache_keep_warm = true
kompressor wrap claude
Here are cost calculations: https://optscale.ai/blog/claude-code-prompt-caching If you run long tasks in Claude Code, how often does the cache expire while a tool is still running? Have you seen this with other providers too?
0 comments