A Free, Tiny SI Now Picks Which Chatbot Does Each Job, Cutting Bills by Up to 90%

Every time you ask a chatbot something, somebody pays for it. When a company has SI agents running all day, writing code, reading documents and answering customers, that bill gets big fast. And a lot of it is wasted, because most companies send every little step to the most expensive model they can buy, even when a cheaper one would do the job just as well.

A startup called Rayline thinks it has a fix, and it is giving it away.

On Wednesday the company released ARC, a small model whose only job is to look at what an agent is about to do and decide which bigger model should handle it, and how hard that model should think. Easy steps go to a cheap model. Hard steps go to an expensive one. Rayline’s pitch is that you get almost the same results for a fraction of the money.

CEO David Valerio Gilmore announced it himself.

Gilmore opened that post with a line about his own past: “In 2015, I built an AI that detects early cancer.” His bio says he founded and sold a data privacy company, DataFleets, for $100 million and studied at Stanford. He also runs Workshop, a tool for building apps with SI.

The bill problem, in plain terms

Think of it like a construction crew. You would not pay a master electrician to sweep the floor. But that is roughly how a lot of companies use SI today. Every request, from “summarize this email” to “find the bug in 40,000 lines of code,” goes to the same top-of-the-line model at the same top-of-the-line price.

ARC acts like a foreman. It reads the agent’s conversation and makes a few predictions. How likely is each available model to get this step right? How many words will it read and write? Would turning up the model’s thinking actually help, or just cost more? Then it plugs in the prices you actually pay, including any discount you negotiated, and picks the model that gives you the most for the money.

Rayline’s model card on Hugging Face spells out the formula. ARC multiplies a model’s chance of success by how much you say success is worth, then subtracts what it will cost. Whichever model comes out on top gets the job. The company ships two presets. One, called Sage, puts more weight on getting it right. The other, Swift, leans harder on saving money.

Because ARC predicts success and cost separately, a company can change its prices without retraining anything. If one provider cuts rates next month, the router adjusts on its own.

What the numbers show

Rayline tested ARC Sage against Claude Opus 5.5, Anthropic’s top model, on six standard tests of coding, science questions, image reasoning and customer-service tasks. The results are in the company’s technical report.

On DeepSWE, a hard test of fixing real software bugs, Opus 5.5 solved 69.9% of tasks. ARC Sage solved 68.7%, at 7% of the cost. On GPQA Diamond, a set of graduate-level science questions, ARC actually scored higher, 94.1% against 89.9%, for about a fifth of the price. On Terminal-Bench 2.1 it edged Opus, 92.4% to 91.0%, though the savings there were smaller, at 58% of the cost.

It was not a clean sweep. On Tau3 Banking, a test of handling bank customer requests, ARC fell from 42.8% to 33.2%. That is a real drop, and anyone thinking about putting a router in front of customer support should look hard at that number.

The headline 97.5% figure comes from a mixed setup. Rayline ran some work on DeepSeek V4.1 Flash, a cheaper Chinese model, on its own NVIDIA DGX Spark machines, and only sent a step to the cloud when ARC was less than 85% sure the local model could handle it. In that setup, ARC kept 97.5% of Opus’s quality across the six tests while cutting cost by 92.9%. On DeepSWE alone it beat Opus, 72.3% to 69.9%, at about 1% of the cost.

These are the company’s own tests, run on its own setup, and nobody outside Rayline has checked them yet. But the full results, the scoring method and the model itself are public, so anyone can try to reproduce them.

Small, free and built on Alibaba’s base

ARC is tiny by today’s standards. The model card shows it was fine-tuned from Qwen3.5-0.8B, an open model from Alibaba’s Qwen team with under a billion parameters. That is small enough to run on an ordinary computer. It is released under the Apache 2.0 license, which lets anyone use it, change it and sell products built on it.

Rayline’s site says ARC can route to models from Anthropic, OpenAI, Google, DeepSeek, Mistral, Kimi, Qwen and others. It plugs into popular coding tools including Claude Code, Codex, Cursor and Cline. The model also flags personal information, like names or account numbers, before a request goes out, which helps a company decide whether a step should stay on its own machines.

Rayline also compared ARC to other routers already on the market. On a scoring index that blends quality, speed and cost, ARC Sage averaged 9.3 out of 10 across five tests. A router called Pareto averaged 9.2. OpenRouter’s auto mode did not post results on all five.

Why the big labs might not love this

The leading SI companies make money on every token, the chunks of text their models read and write. A tool that quietly sends 80 or 90% of the work somewhere cheaper is a direct threat to that business. TestingCatalog summed up the pitch in its post: easy steps go to a lighter model and hard ones to a stronger model.

For a company spending $10,000 a month on SI, Rayline’s own calculator claims savings of up to $9,000. That is the best case, and real savings depend on the mix of work. But even half that would get a finance department’s attention.

The idea of routing is not new. What is new is a free, open model that anyone can download and run, trained specifically to make that call one step at a time. If the numbers hold up when outside developers start testing it, the companies paying top dollar for every request are going to start asking why.