Updated October 2026

Measured on real traffic.

How much of a Claude Code bill Probe0 and other cost tools save, on the same traffic.
of the bill saved by default

12.8%

of the bill saved by default

No model swaps. Nothing to set up.

with model step-down on Heavy

41.8%

with model step-down on Heavy

Easy turns go to a cheaper model.

live answers judged blind

314

live answers judged blind

To check the savings don't cost quality.

Share of the bill saved

Probe0 shrinks tool output without disturbing the part of the prompt your provider already has cached, and stops paying for cache time nobody uses. SuperCompress adds a running summary after tool calls, and that text is cached and re-read on every turn after it.

  • Probe012.8%
  • RTK9.6%
  • Headroom (claimed)5.2%
  • caveman0.3%
  • Claude Code compaction<0.1%
  • SuperCompress28.5% more

With model step-down on

Step-down sends easy turns to a cheaper model in the same family. It's off by default. Turned on, the saving grows with how much traffic you let it move.

  • Heavy41.8%
  • Medium36.2%
  • Light24.0%
  • Off12.8%
  • RTK, for reference9.6%

Does it cost quality?

Every saving that changes what the model sees was tested live on real models, with objective checks first and a blind judge after.

  • Opus → Sonnet on easy and hard turns

    25 prompts, 3 models, blind-judged

    Same objective pass rate. Judge gaps smaller than the judges' disagreement with each other.

    On by default

  • Lower effort on tool turns

    25 prompts × 3 effort levels

    17 of 17 objective checks passed at every level.

    On, with a floor

  • Folding long shell output

    12 prompts of real command output

    About 1% expected answer cost. When a detail was folded away, the agent asked to re-run the command instead of guessing.

    On by default

  • Cache timing and fan-out

    Not needed

    Changes no byte of the prompt, so it can't change an answer. Worth 7% of the bill on its own.

    No risk

Compressing long documents

Probe0 scores each line against the question being asked and keeps the ones that name what it's about, so the answer survives a deeper cut.

MeasureProbe0SuperCompress
Answers kept+0.8 pts ahead97.8% (better)97.0%
Tokens removed+2.6–3.0 pts ahead67.6–67.8% (better)64.6–65.2%
Suite runs that pass the 98% / 55% bar2× as many ahead8 of 16 (better)4 of 16
Lowest average removal in a run+5.6 pts ahead58.0% (better)52.4%

How it was measured

Twelve days of one developer's real Claude Code traffic, replayed through each tool on the same bill. Quality checks ran live on real models.

Try it on your own agents.Free beta for Mac.

Download for Mac