Route every task to the cheapest model that can actually do it
Frontier models get handed trivial work every day, and the bill never notices. Arbiter routes each task against the capability, cost, and safety policy you set, and absorbs the coordination work on a local brain that costs nothing per call.
There is a quiet tax in most AI operations, and it is paid one task at a time. A frontier model, priced for hard reasoning, gets handed something trivial: sort a list, classify a string, decide which of two paths a job should take. It answers correctly, because of course it does. It also charges frontier rates for kindergarten work, and nothing in the stack flinches. Do that a few thousand times a day and the waste stops being rounding error.
The fix is not a cheaper model for everything. Some work genuinely needs the strongest model you can reach. The fix is matching each task to the cheapest model that can actually do it, and having something in the loop that knows the difference.
Sufficient, not maximal
Arbiter gives operators a place to make the capability, cost, and safety tradeoff explicit, then routes dispatched work against it. The principle is sufficiency, not maximum horsepower on every call. Easy work goes to cheap capacity. Hard work earns the expensive tier. The route each task took is recorded with the reasons behind it, so the decision is auditable after the fact rather than a black box. The learned cost optimization on top of this is still being hardened for release. The policy you configure is what drives routing today.
A tier that costs nothing per call
The cheapest tier of all is the one that never places a metered call. Arbiter hosts its own model on your GPU, and that local brain absorbs the coordination work an operation generates all day long: triage, classification, short summaries, the ops chat, deciding where a task should go. That work is real inference, and on most stacks it quietly runs up a bill of its own. Here it runs on hardware you already own, in the decision path, with no per-token charge attached.
To be honest about it, local is not free in the absolute sense. You still pay for the card, the power, and whatever providers you choose to connect for the heavy work. What you stop paying is a frontier rate on every trivial decision an operation makes just to keep itself moving.
Cheap is only safe next to a guardrail
Optimizing for cost alone is how a runaway job quietly empties a budget. Cost policy only works when it is paired with safety policy, so Arbiter keeps route constraints, budget ceilings, and approvals for sensitive work right beside the cost controls. That pairing is the subject of Guardrails for Autonomous Agents, and it is the reason cheap-by-default does not have to mean reckless-by-default.
The conductor of a sovereign stack
Arbiter shares this cost-aware routing vocabulary with Warden, which executes work across providers, and it shares the local-brain principle with Curator, which keeps memory on hardware you own. The whole design is laid out in The Sovereign AI Engineering Stack.
Arbiter is a proprietary working alpha in public binary release preparation. Follow the public release channel or RSS feed for updates.