Rayline launches ARC, an open-weight model that routes agent steps between models

Rayline, founded by David Gilmore, has released ARC, which it calls a unified model for routing AI. For each step of an agent's work, ARC picks a model and how much it should think, using the conversation so far, the available compute and cached context. It can switch models, change the reasoning level or stay put, and can mix cloud and local models in one session.

Rayline claims ARC reaches 97.5% of Claude Opus 5.5's quality at up to a tenth of the cost. In a Hermes demo, it says the same task cost 49% fewer tokens than running Opus 5.5 on every step. A recap of its technical report puts DeepSWE at 68.7% against 69.9% for Opus 5.5, at 7% of the cost per task. These are Rayline's figures.

The 800M-parameter router is open-weight under Apache 2.0 on Hugging Face (rayline-ai/ARC-1.0) and can run locally. Rayline offers it three ways:

• Generation API: Rayline routes and generates. The ARC Sage preset favors quality and ARC Swift favors efficiency, each priced like a model. ARC Dynamic lets you tune the cost and quality tradeoff and charges the selected model's rates with no markup. It works as a drop-in model in Claude Code, OpenCode, Cursor, OpenClaw and other agent harnesses.
• Routes API: returns a model and reasoning level, and the application makes the inference call on its own infrastructure.
• Decisions API: returns choices, probabilities and scores for any input, to feed routing policies, gate expensive calls or prioritize queues.