We benchmarked Tempest - 86% fewer tokens on real repos!

We promised we would improve and publish Tempest's token efficiency results soon. Here we are.

7 open-source repos. Same agent, same model, same prompts. The only variable was whether Tempest's Token Intelligence is wired in.

Median results across the board:

• Upto 88% fewer tokens

• Upto 92% fewer tool calls

Your agents stop re-reading the same codebase from scratch on every run. One shared knowledge graph, indexed once, used by every parallel session.

Full methodology and reproduction steps are in the repo. Nothing is hidden. You can run it too.

Github Link: github.com/tempestai-dev/tempest