Jevgrep cuts coding agent token cost by 26% on SWE-bench
On the same 10 SWE-bench tasks, the same coding agent cost $7.62 without Jevgrep and about $5.66 with it, with the solve rate unchanged at 8/10.
The same coding agent & the same 10 SWE-bench tasks. ✅ Without Jevgrep: 8/10 solved — $7.62 ✅ With Jevgrep: 8/10 solved — ~$5.66 There is about a 26% cost savings for the same solve rate. So coding agents can burn a lot of tokens just wandering around an unfamiliar repo trying to figure out which files matter. Jevgrep gives that job to Jev instead. This is a perfect use for a Decision Model. If you want to know where authentication is checked, jg searches the repo semantically and returns the relevant: 📁 files 📍 line numbers 📄 source excerpts 🔎 reading leads Then Claude Code/Codex does the actual coding. Instead of spending expensive coding model's context figuring out where to look... ...give it the relevant source first. In their 10-task Python SWE-bench test: 💰 Baseline: $7.62 💰 Jevgrep + estimated Jev cost: ~$5.66 🎯 Both: 8/10 solved ⚠️ This is not evidence that every coding job will be 26% cheaper. Jevgrep is MIT-licensed. 🔗 Link in ALT.