Back home

ARC / Rayline ARC

x+web 0.93
Visit source

Intro

ARC / Rayline ARC is a routing decision model announced by Rayline on November 11, 2026. Founder David Valerio Gilmore wrote in the announcement post at 2026-11-11T18:26:29Z: “Today, we’re introducing ARC by Rayline: the world’s first unified model for routing AI.” Officially it is also called the Adaptive Routing Controller: it reads an agent’s conversation so far and produces signals for choosing a model and a reasoning level for the next step. The product page is rayline.ai/arc, the site is rayline.ai, the technical report is ARC: Learned Model Routing for Long-Horizon Agentic Workflows, the docs are at docs.rayline.ai/arc, and the weights are at huggingface.co/rayline-ai/ARC-1.0.

Officially, ARC separates predictions from policy. It predicts success and token usage; the application supplies prices and decides how to balance quality, cost, and local execution. Updating prices does not require retraining those predictors when the model and serving configuration stay fixed. The model card lists signals for: predicted probability that an eligible model will complete a task; estimates of uncached input, cache reads, cache writes, and output; the expected benefit and cost of supported reasoning settings within a model; PII detection in the supplied input; classification of requested operations and domains; and a general-purpose question head that takes a question and answer options (for example, “Does this request need a stronger model?”).

There are two official ways to use it. The Generation API lets Rayline handle routing and generation. The Routes API returns a routing decision for you to run on your own endpoints. The Generation API presets are configurations of the same router. In a follow-up post, the company introduced ARC Sage, which puts more weight on quality for demanding work, and ARC Swift, which focuses on efficiency for everyday tasks; both are priced like models, with fixed input, output, and cache rates. ARC Dynamic lets you tune the cost/quality tradeoff and, officially, you pay the selected model’s inference rates with 0% markup. The Hugging Face model card gives the utility rule Utility = V × predicted success − expected cost, with V = 4.5 for Sage and V = 0.1 for Swift, plus a prefer-local policy that uses an eligible local model when it clears a chosen predicted-success threshold and otherwise falls back. Officially, you can use ARC as a drop-in model in Claude Code, OpenCode, Cursor, OpenClaw, and other agent harnesses.

On the model itself, a follow-up post says ARC was trained for routing across multi-turn conversations. It uses the conversation so far, available compute, and cached context to choose the next model and reasoning level; it can switch models, adjust reasoning, stay with the current model, or mix cloud and local models in one session. Officially, Rayline updates ARC as new models arrive. The router is an 800-million-parameter model that can run locally, with open weights under Apache 2.0, fine-tuned from Qwen3.5-0.8B. At the time of this scan, rayline-ai/ARC-1.0 was a public Hugging Face repository. The docs Availability page still listed “ARC weights and model-specific self-hosting instructions” as coming soon, so check that page and the model card for the current status before you rely on self-hosting details.

Quality and cost figures are all from official materials, not independent tests by this site. The announcement post states “97.5% of Opus 5.5 quality” and “Up to 1/10th of the cost.” The technical report and model card say that, in the paper’s hybrid edge–cloud configuration, across six benchmarks, it retains 97.5% of Opus 5.5 quality at 92.9% lower cost, using DeepSeek V4.1 Flash and a multiple-DGX-Spark setup with a 0.85 prefer-local threshold. The homepage also shows another comparison: 97.4% quality retained, ARC at 14.3% of cost, and ARC + local at 7.1%. The model card also reports official per-benchmark examples: on DeepSWE, Opus 5.5 at 69.9% success, ARC Sage at 68.7%, at 7% of baseline cost; on Terminal-Bench 2.1, Opus 5.5 at 91.0%, ARC Sage at 92.4%, at 58% of baseline cost.

Key points:

Explain

Why it matters

This is a formal product name Rayline’s founder announced on November 11, 2026, with a product page, a technical report, documentation, and public Hugging Face weights, not a rumor. It is aimed at the routing problem in agent workflows—paying frontier-model prices on every step—and ships an 800-million-parameter open-weight router that can run locally, plus hosted Generation and Routes APIs. For anyone tracking model routing, agent inference cost, or open-weight decision models, this is a new name that deserves its own entry. The previous 04:32 scan did not record the announcement; this slot adds it.

This page only restates public official materials. The announcement’s “world’s first” phrasing, the 97.5% quality and “up to one-tenth the cost” claims, and the slightly different comparison figures on the report, the model card, and the homepage all come from Rayline; this page does not independently evaluate them.

Limits

Quality and cost figures are company-reported on Rayline’s chosen benchmarks, model pool, and hybrid edge–cloud setup, and the homepage figures are not identical to the report’s. The docs Availability page still listed some self-hosting instructions as coming soon; rely on the latest official docs and model card. Do not treat this entry as a claim that ARC is already the default in every agent tool, or treat official marketing claims as third-party benchmark results.

Usage

  1. Read the product page at https://rayline.ai/arc , and compare it with the site at https://rayline.ai/ , the technical report at https://rayline.ai/research/report , the docs at https://docs.rayline.ai/arc , the weights at https://huggingface.co/rayline-ai/ARC-1.0 , and the announcement at https://x.com/davidvgilmore/status/2120583960447029699 ;
  2. Follow the official docs to choose the Generation API (the ARC Sage / ARC Swift presets, or tunable ARC Dynamic) or the Routes API (you run inference); for self-hosting, use the latest model card and the docs Availability page ;
  3. When you cite 97.5% quality, cost reductions, “world’s first,” or per-benchmark figures, attribute them to the 2026-11-11 official announcement, the technical report, the site, or the model card, and re-check the upstream source; the homepage figures are not identical to the report’s ;
  4. Do not describe ARC / Rayline ARC as a new version of a frontier lab model, or confuse it with other projects that also use the name ARC; this page only restates official materials, so re-check the upstream source before you cite it elsewhere.

Advertisement