Benzi: compiler-backed coding agent hits 82.4% SWE-bench Verified for under $13
DonkeyTheKing · reddit · 2026-10-24
Benzi, a new free-beta coding agent (local compilation, model-agnostic), scores 82.4% on SWE-bench Verified while claiming to be 2x faster, cheaper, and reading 2-6x less source code than mainstream agents.
The core idea: current agents discover program structure by repeatedly searching and reading code or retrieving approximate embedding chunks. Benzi instead compiles the repo into a resolved, queryable map of symbols, calls, references, inheritance, and data flow, exposing deterministic tools like "what calls this?", "where did this value come from?" and "what will this edit affect?" — the model still reads source when needed but never has to rediscover codebase structure from text.
It tracks three truth tiers — RESOLVED (statically proven), CANDIDATE (static analysis can't prove the edge), OBSERVED (what actually happened at runtime) — so it uses deterministic intelligence without pretending static analysis is infallible. Supports Python, JS/TS, Java, C#, C/C++, Go, Rust, Ruby, plus deterministic HTML/CSS/DOM resolution.
More from coding & agent
- Polylane Launches for Cloudflare: Proactive Agents Watch Workers, Merge ~8 PRs a Day — dinasaur_404 · 2026-10-24
- Factory tested Reflection's Beam pre-launch via Together AI for day-0 support — togethercompute · 2026-10-24
- Kimi K3.1 vs GPT-6 Astra on the same agentic 3D task: full token and cost breakdown — GapNew4766 · 2026-10-24
- Bench shows GPT-5.6/6 Luna crush multi-agent coordination while DeepSeek V4 fails — JoshPurtell · 2026-10-24
- Payman AI launches MCP server for natural-language payments and payee management — modelcontextprotocol · 2026-10-24
- Grok Bot adds secure vault so agents can log in and pay without seeing your secrets — XFreeze · 2026-10-24