Havrin moves the routine steps of a Claude Code session to cheaper models and your own machine. Claude keeps the hard parts. Your tests check every result.
Searches and file exploration that Claude hands to a subagent run on Sonnet instead of Opus. Built.
Long tool output is cut before it enters the conversation, because Claude rereads all of it every turn. Testing.
On a machine with a capable GPU, a local model drafts routine fixes and your tests decide what is kept. Research.
of cost was rereading cached context, which happens on every turn.
of cost came from subagents, all of them on Opus. Havrin moves these first.
tokens of context on an average request, across 10,915 requests.
30 days of one developer's sessions, weighted by cost
The meter replays your last 30 days from local Claude Code logs and estimates what each change would have saved. Nothing leaves your machine.
$ havrin meter --days 30 --what-if cost by model opus 100% sonnet 0% subagents 48% of cost what-if routine subagents on Sonnet up to 19% context capped at 150k up to 19% both up to 38%
One email when it opens. Every change ships only if it matches plain Claude Code's pass rate, and the benchmark is published first.
Your email is used only for beta news and never shared.