Offloop says its 3B dispatcher beats Codex and Claude Code on GDPval at $1.65 a task
dr_cintas · x · 2026-07-24
Offloop says its new multi-agent system beats Codex and Claude Code on real-work benchmarks.
According to the post, the system uses a 3B-parameter dispatcher called D1 to decide which agent should act next — and, just as importantly, which agents should stay quiet. On the same 220 GDPval tasks, it reports costs of $1.65/task versus $5.20/task for Codex and $14.38/task for Claude Code. The team says the project was built by people who left Manus.
Related event: Offloop's Multi-Agent Orchestration Framework Tops GDPval Benchmark(6 posts)→
More from coding & agent
- claude-tap exposes live traces for Claude Code and Codex-style agents — dotey · 2026-07-24
- A local pipeline turns text or images into animated 3D assets in under 30 minutes — Tamerygo · 2026-07-24
- ByteDance-linked verl framework gets a SGLang contributor as Agentic RL talk lands — VoidAsuka · 2026-07-24
- Rethinking Agent Workspaces: Why Ephemeral Environments Beat Durable Ones — generativist · 2026-07-24
- Fable ultracode runs for weeks on a Mac mini to build cross-platform AI autocomplete — josh_bickett · 2026-07-24
- Open-Source AI Hedge Fund Project Adds Backtesting and Benchmarking — virattt · 2026-07-24