Skip to main content

​
When free models work

Routine work — classifying a prompt, drafting a plan, verifying a result — runs on free local models. They’re fast, cost nothing per call, and run on your own hardware. For reasoning-heavy classification, the platform uses thinking-tuned local models.

​
When paid models earn their cost

Architecture work, implementation, and high-risk decisions go to paid models — Claude, OpenAI, and other OpenAI-compatible endpoints. Routing is policy-driven; you decide which categories of work go where.

​
Fallback that just works

Every paid provider has a health score. When one fails — rate limit, outage, malformed response — the platform de-prefers it for subsequent jobs and falls through to the next provider on your list. You don’t need to write retry code; the platform handles it.

​
Per-call cost tracking

Every paid call is recorded with model, token counts, and cost. You can see what each job cost, broken down by model and step. The dashboard surfaces this per project so you know exactly where spend is going.