Offloop's Multi-Agent Framework Tops GDPval Benchmark

Startup Offloop has released a multi-agent orchestration framework that claims to outperform Claude Code and Codex on the GDPval benchmark, which covers knowledge tasks tied to $2.4 trillion in U.S. economic activity. The system emphasizes intelligent scheduling and cost control, utilizing a router with only 3B active parameters. This development highlights a pivotal shift in AI: true capability is moving from individual models to external harnesses and orchestration systems, demanding attention from developers and enterprise users alike.

Confirmed

Based on the provided posts, the Offloop framework has the following core features and data points:

Intelligent Scheduling: Upon receiving an instruction, the system automatically activates suitable agents, allocates tasks, performs quality checks, and requests a single human approval only at the final step. The core of this is a scheduler named D1, which decides the next action and coordinates the overall workflow.
Benchmark Performance: Achieved SOTA on the GDPval benchmark, which involves generating documents, spreadsheets, and slides, surpassing both Claude Code and Codex.
Cost and Parameters: The orchestration router D1 uses 3B active parameters. It can scale to more agents without letting compute costs spiral out of control, reportedly achieving SOTA on routing benchmarks at an 80% lower cost.
Team Size: The system was reportedly developed by a team of only 4 people.

Why it matters

The current trend in the AI agent space is shifting from simply enhancing individual model capabilities toward more complex multi-agent collaboration and orchestration. Offloop demonstrates the feasibility of achieving highly effective workflow results while drastically reducing computational costs through intelligent routing and task allocation, offering a novel solution for efficient, enterprise-grade knowledge work automation.

2026-07-23 ~ 2026-07-24 · 10 related posts

Primary sources

Offloop claims a multi-agent harness beat Claude Code and Codex on GDPval — teortaxesTex · 2026-07-23 23:54 (UTC+08:00)
Offloop says its 3B dispatcher beats Codex and Claude Code on GDPval at $1.65 a task — dr_cintas · 2026-07-24 02:41 (UTC+08:00)
Offloop says its multi-agent router delivers SOTA at 80% lower task cost — umesh_ai · 2026-07-23 23:12 (UTC+08:00)

[source] Offloop says its multi-agent router delivers SOTA at 80% lower task cost — umesh_ai · 2026-07-23
[source] Offloop claims a multi-agent harness beat Claude Code and Codex on GDPval — teortaxesTex · 2026-07-23
Offloop says its four-person agent harness beat Claude Code and Codex on GDPval — hey_abusiddik · 2026-07-24
Offloop’s multi-agent harness beats Claude Code and Codex on workflow benchmarks — SucceededMind · 2026-07-24
[source] Offloop says its 3B dispatcher beats Codex and Claude Code on GDPval at $1.65 a task — dr_cintas · 2026-07-24
Offloop says its multi-agent harness tops GDPval, ahead of Claude Code and Codex — azed_ai · 2026-07-24
Offloop says its 4-person multi-agent harness beat Claude Code and Codex on GDPval — heyshrutimishra · 2026-07-24
Offloop says its 4-person multi-agent harness beat Claude Code and Codex on GDPval — rohanpaul_ai · 2026-07-24

2 near-duplicate retellings: iamfakhrealam · SimplyAnnisa