Cut agent costswithout gamblingon quality
Tokenslash finds high-cost steps in traces you already collect, replays them with cheaper models, and every swap against your own evals — then hands you a cost map and a playbook for the next run.
scoring against your evals…
Replaying agent ci-test-fixer on fix-ci-flake. Steps swap from opus-4 to cheaper models when evals pass.
How it works
We sample expensive tasks from your traces that appear repetitive. Then we test the same tasks with cheaper models.
01
We watch real runs
Tokenslash reads the traces your AI agents already produce. Every step, every model call, every cost.
02
We re-run them cheaper
We replay the same tasks with a cheaper model swapped in at each step, one step at a time.
03
You choose what to keep
Every result is checked against your own tests. You (or your agent) decide if it's worth it to swap.
Measured, not guessed
We do not ask you to trust a cheaper model. We ask you to trust your own tests.
Without us, downgrading is a gamble
Most tasks do not need a frontier model. But nobody measures the quality cost of each swap, so every downgrade is a risky guess. We replace the guess with a number.
Built for coding agents first
Code is verifiable. Tests pass or they do not, builds compile or they do not. That means quality is measured by your own harness, not judged solely by another model.
Applying a swap is one click
No rewrites, no new framework. The model at each step is just a config value. We tell you which values are safe to change and which ones result in quality drops.
Don't want to make the decision? Your agent can call our API and make it for you.
Custom use case? Plug in your own eval
Not every agent writes code. Bring the checks that already define success for your workflow, and Tokenslash will gate every model swap against them the same way.
Granular cost control
Control exactly which model executes which step, ensuring you only pay for frontier intelligence when you actually need it.
agent: ci-test-fixer steps: parse_issue: model: opus-4 haiku-4.5 locate_files: model: opus-4 haiku-4.5 retrieve_context: model: opus-4 sonnet-4 draft_plan: model: opus-4 sonnet-4 write_patch: model: opus-4 # kept because quality drops on cheaper models run_tests: model: opus-4 glm-4.5 open_pr: model: opus-4 haiku-4.5 # verified by tokenslash: saved 82%, 0 quality regressions
| agent | ci-test-fixer |
|---|---|
| before | $5.29 |
| after | $0.96 |
| saved | $4.33 (82%) |
| steps swapped | 6 / 7 |
| steps retained | 1 / 7 |
| best swap | run_tests |
| biggest savings | $2.82 (98%) |
| eval pass rate | 100% |
| quality regressions | 0 |