How Hearth cut their LLM bill 85% in three days.

Every workload routed to the same frontier model. The classifier needed nothing close to its reasoning depth. The drafting tasks didn’t need its context window. They were renting capability they never touched, three times over.

Hearth handed over 312 anonymized prompts across the three task types, sample-weighted to match their production traffic. We ran each task type as a separate sub-audit so we could recommend a model per workload, not one blended choice.

• 312anonymized prompts collected
• 4,804candidate models filtered
• 41passed hard constraints (US, JSON, tools, 32k)
• 8finalists shortlisted per task
• 2,496inference calls + 750 judge calls
• 2h 18mwall-clock end-to-end (10× parallel)

Hearth flipped the switch in a single 4-hour sprint. They put OpenRouter in front of every model call, added three routing rules keyed off their internal task_kind tag, and shipped behind a 10% canary for 48 hours before fully rolling out. Total integration: 72 hours from receipt to production.

route(task_kind) {
  case "classify_ticket":
    return "alibaba/qwen3-32b"
  case "summarize_thread":
    return "deepseek/v3.1"
  case "draft_reply":
    return "anthropic/claude-haiku-4.5"
  default:
    return "openai/gpt-5"  // fallback
}