Skip to content

Model routing estimator

Send each task to the model it needs.

An illustrative estimator for routing small tasks to small models and hard tasks to frontier models, in relative units. Not a quote, a price list or a forecast.

Illustration: the glass Confluence ring sends a beam of tasks into a glass prism, which splits it into three thin strands to a small glass cube, the small models, and one thick strand to a large glass orb, the frontier models.

Small tasks to small models. Hard tasks to frontier models.

Some agent work is small: renames, tests for simple functions, formatting and summaries. Routing can send that work to a small or local model and save the frontier model for the tasks that need it. This page shows the math with your own numbers.

Illustrative estimator

Change the inputs to see how the share of hard tasks and the cost gap between models move the total. These are relative units, not prices.

Frontier model for everything: 200 units a day.Routed by difficulty: 57.5 units a day.Illustrative reduction: 71 percent.

Illustrative only. One unit is the cost of one task on a frontier model. Values outside the allowed range are clamped. Live updates need JavaScript; the worked example below shows the same math.

Worked example

Say a fleet runs 200 tasks a day, 25 percent of them are hard, and a small model costs 5 percent of a frontier model per task.

  • Frontier model for everything: 200 tasks at 1 unit each, so 200 units a day.
  • Routed: 50 hard tasks at 1 unit each, plus 150 small tasks at 0.05 units each, so 50 plus 7.5, or 57.5 units a day.
  • Illustrative reduction: about 71 percent.

Real results depend on tokens per task, retries, verification passes and each vendor's own pricing. Treat this as a way to reason about routing, not a forecast.

How OmniConflux routes

In the design, you set the rules: which vendors and models each kind of task may use, and whether local models come first. The coordinator is designed to place each task on a model and a machine, and to track rate limits across the fleet. See the fleet architecture guide for details.

Join the waitlist

Questions

Are these real prices?
No. The estimator uses relative units, where one task on a frontier model costs one unit. It is an illustration of the routing math, not a quote or a price list.
What counts as a hard task?
Work that needs deep reasoning across a large codebase, such as a design change or a tricky bug. Formatting, renames, boilerplate and summaries are usually small tasks.
Does routing lower quality?
It can if a task is sent to a model that is too small. That is why every task, small or hard, still goes to a verifier from a different vendor, and work that fails comes back for rework.
Which models can I route between?
By design, any you bring: a vendor's own tool on your machine under your login, your own API key, or a local model through Ollama or llama.cpp.

All product names are trademarks of their owners.