Cost and models

Why cost matters here

An ensemble can run multiple live RouteKit models plus judge/synthesis calls on every prompt. Fusion trades extra calls for better answers, so usage can add up faster than a single-model RouteKit session.

For provider selection, model ids, local MLX cache commands, and mixed panels, read Models and panels. This page focuses on spend control.

Ways to control cost

Pick cheaper models

For an embedded project router, choose cheaper namespaced entries from the provider's catalog and use fusionkit doctor to validate the configured IDs. Use routekit models list to inspect available entries only when router.url targets the standalone singleton, then compose those IDs:

fusionkit ensemble add economical \
  --member openai/gpt-5-mini \
  --member openrouter/qwen/qwen3-coder \
  --judge openai/gpt-5-mini

Or commit it via the .fusionkit/ folder.

Shrink the panel

A panel can be a single model plus a judge if you want fusion's synthesis step without a wide fan-out.

Cap the session

Use --budget to stop a session after gateway-observed spend crosses the cap.

fusionkit codex --budget 5

Persist the same cap in .fusionkit/fusion.json:

{
  "version": "fusionkit.fusion.v4",
  "router": { "config": ".routekit/router.yaml" },
  "ensembles": {
    "default": {
      "members": ["openai/gpt-5-mini"],
      "judge": "openai/gpt-5-mini"
    }
  },
  "budgetUsd": 5
}

Preview before you spend

fusionkit config show

shows the effective namespaced-model ensembles and judge.

Inspect accumulated cost later:

fusionkit sessions
fusionkit sessions show <id>

Provider access and pricing metadata are RouteKit responsibilities. FusionKit meters usage returned across that boundary.