Four stacked tools cut Claude Code bills over 90%

Claude Code users are chaining four separate token-reduction tools to cut their bills more than 90%, per a Codepointer breakdown.

That matters because for anyone running agents on a subscription, the real ceiling isn't the invoice. It's the usage limit, and every wasted token on a build-log dump is throughput you don't get back.

The stack follows an order. A model router picks the cheapest model for the job first.

Then rtk compresses tool and Bash output going back into context, cutting 60 to 90% off that class of token.

Next comes Caveman, a one-line Claude Code skill that forces terse output. It cut an average 65% of output tokens across ten benchmark prompts while keeping the substance.

Last in the chain is the Headroom proxy. It compresses tool and RAG payloads at the network layer, claiming 60 to 95% reduction.

Because each tool hits a different token class, input, tool output, model output, the savings multiply instead of overlapping.

Firecrawl ran its own 12-technique benchmark separately. Model routing, sending subagent work to Haiku instead of Opus, and path-scoped rules delivered the biggest wins. One prompt-shaping technique hit 91.9% context reduction with no measured quality loss.

The discipline that makes this work is measurement, not stacking. Run ccusage for a week to get a real baseline, then add one tool at a time. Stack all four blind and you'll never know which one is carrying your savings, or which one quietly wrecked your output.

The 90% number will get repeated more than the measurement discipline that makes it safe. Most people stacking these four tools blind won't notice which one degraded their output until the code breaks.

The timing lines up with the margin squeeze on providers. As their per-token profit shrinks, the same compression logic is showing up on the buyer's side of the ledger too.