How to reduce LLM bills effortlessly?

I’m guessing that, like me, you’ve already seen your Claude Code or Cursor bill blow past the planned budget… just because you left the tool running on a complex task for a few minutes (or hours).

Going forward, this problem tends to get worse as AI tools become embedded everywhere in our workflows: writing code, exploring a project, finding an info in a folder… you name it.

And on top of that, with the rising cost of tokens on the latest models, AI spend is becoming a real FinOps concern for the coming months (or even years).

So yes—but you might tell me: what can I do about it, at my scale?

The solution is simple: try to reduce as much as possible the number of tokens sent to the LLM, without degrading the context (or the useful information).

For that, I found a pretty “magic” solution as I’m writing this article 👇

It’s called “Rust Token Killer” or rtk ✨

Rust Token Killer (or rtk for close friends) is an open-source tool developed and maintained by French folks (cocorico 🇫🇷) that acts as a proxy for the various commands an agent / AI IDE often runs to get its work done.

Instead of calling the baseline commands provided by your system (like ls or find) — designed to be read by humans — rtk “wraps” them and replaces the output with optimized, less verbose alternatives, while minimizing information loss.

So forget ls (~150 tokens)

$ ls .
total 24
drwxr-xr-x@  4 myastr0  staff   128  9 mar 08:50 assets/
drwxr-xr-x@ 17 myastr0  staff   544  9 mar 18:19 components/
-rw-r--r--@  1 myastr0  staff  1338  9 mar 18:38 config.ts
-rw-r--r--@  1 myastr0  staff  1001  9 mar 19:31 constants.ts
-rw-r--r--@  1 myastr0  staff   847  9 mar 18:34 content.config.ts
drwxr-xr-x@  3 myastr0  staff    96  9 mar 18:34 data/
drwxr-xr-x@  9 myastr0  staff   288  9 mar 08:50 i18n/
drwxr-xr-x@  5 myastr0  staff   160  9 mar 18:34 layouts/
drwxr-xr-x@  5 myastr0  staff   160  9 mar 08:50 pages/
drwxr-xr-x@  4 myastr0  staff   128  9 mar 08:50 styles/
drwxr-xr-x@ 13 myastr0  staff   416  9 mar 18:34 utils

and say hello to rtk ls (#~35 tokens -80%)

$ rtk ls .
assets/
components/
data/
i18n/
layouts/
pages/
styles/
utils/
config.ts  1.3K
constants.ts  1001B
content.config.ts  847B

📊 3 files, 8 dirs (3 .ts)

The outputs produced by rtk are much less verbose, but they remain usable by an LLM for classic tasks like file exploration and context enrichment.

Cool, it prints less text… but how do you measure what you’re saving? 📊

Where rtk really won me over is tracking. With rtk gain, I get a detailed report of what I’m saving (in tokens).

Below is my report from last week 👇

$ rtk gain


📊 RTK Token Savings
════════════════════════════════════════

Total commands:    2,927
Input tokens:      11.6M
Output tokens:     1.4M
Tokens saved:      10.3M (89.2%)

By Command:
────────────────────────────────────────
Command               Count      Saved     Avg%
rtk find                324       6.8M    78.3%
rtk git status          215       1.4M    80.8%
rtk grep                227     786.7K    49.5%
rtk cargo test           16      50.1K    91.8%

If this report isn’t convincing enough, here’s a breakdown based on the official Anthropic API pricing (Claude, March 2026).

What that means in real money 💰

The 10,3 M tokens saved by rtk correspond here to input context. Using the per-million-token prices ((MTok)) from the Anthropic documentation for recent models:

• For Opus 4.6: ~51,50 $ saved over the week
• For Sonnet 4.6: ~30,90 $ saved over the week

Annualized (× 52), that’s about 2 680 $ saved per year with Opus 4.6 — the number speaks for itself.

So, without changing model or provider, with rtk you can aim for ~1 600 $ to ~2 700 $ saved per year (order of magnitude, depending on the model) on a single machine. At the scale of a team (e.g., 30 devs), the order of magnitude becomes significant quickly.

So, ready to save a ton of money? 🚀

For that, nothing simpler: install and configure rtk (I’ll let you follow the official README), and you’re ready to watch your bills go down! 😉

Thanks for reading this far 🙇‍♂️

If you have feedback on this article (or if you liked it), feel free to message me on Bluesky, share it on LinkedIn, or elsewhere.