Hacker News
Show HN: Dirac Agent Tops TerminalBench-2 at 65.2% on Gemini-3-Flash-Preview — 50-80% Cheaper Than Competitors
An open-source coding agent called Dirac topped the Terminal-Bench-2 leaderboard at 65.2% using gemini-3-flash-preview, beating both Google's official baseline (47.6%) and top closed-source agent Junie CLI (64.3%). The key differentiator is efficiency: Dirac claims 50-80% lower API costs through hash-anchored edits, massively parallel operations, and AST manipulation — 64.8% cheaper than competitors (2.8x cost reduction). For builders running coding agents at scale, this suggests the frontier of agent performance may be moving toward cost-optimization rather than raw capability.
↳ Follow the thread