Benzi coding agent hits 78.2% SWE-bench reading far less code than Claude Code

DonkeyTheKing · reddit · 2026-09-16

A new open-source coding harness called Benzi takes a contrarian approach: instead of stuffing source code into context, it gives the model deterministic intelligence via tool calls.

Core idea: Current agents rely on retrieved code snippets or high-dimensional embeddings, which balloon token counts, cause context drift, and break down on multi-file refactors when line numbers shift. Benzi's compiler informs the model of the blast radius before and after edits, with full static analysis — often eliminating the need to read code at all.
Benchmarks: Benzi Sonnet read 9,125 lines to complete the same tasks vs. Claude Code at 20,704, DeepSeek's harness at 43,598, and OpenCode at 65K+ (disqualified for repeated failures). It scores 78.2% on SWE-bench Verified at under 10¢ per fix (using V4flash).
Extras: Clear "truth tiers" separating static-analysis-verifiable facts from runtime-checked ones, a runtime tracer, self-aware mid-task model upgrades, and support for 10 languages including Python, TypeScript, Go, and Rust.
The author argues that while the industry pours money into bigger context windows, indexing a repo beats reading raw source on both cost and comprehension. Open source, with a live demo and benchmarks available.

Original post →

Benzi coding agent hits 78.2% SWE-bench reading far less code than Claude Code

More from coding & agent