Multi-Model AI Routing with OpenRouter — Cost Optimization Strategy

Hiramise runs multiple AI workloads: layout generation, catalog image analysis, compliance checks. Using the same model for everything causes costs to spike. Here's how we split models by task type using OpenRouter.

Why OpenRouter

OpenRouter provides a single OpenAI-compatible API endpoint for dozens of models. Switching providers requires only a model ID change — no code changes.

• Single API key manages Anthropic, Google, Meta, and more
• Fallback routing — automatic failover when the primary model is overloaded
• Built-in per-token cost monitoring dashboard

Task-Based Model Routing Strategy

All AI calls pass through a single client wrapper (app/ai/client.py). Model IDs are isolated in a constants file (app/ai/models.py) to minimize switching costs.

• Layout generation (complex reasoning): claude-opus-4.7 — accuracy first
• Catalog image analysis (Vision): claude-sonnet-4.6 — speed + cost balance
• Simple text summarization/tagging: claude-haiku (lowest cost)
• Compliance checks (long cacheable context): Opus + prompt caching

Cost Optimization Results

Compared to routing everything through Opus, monthly AI spend dropped by roughly 40%. Maintaining Opus for layout generation while switching the frequent Vision calls to Sonnet was the key lever.

• Vision analysis calls: 65% of all calls → up to 5× cost reduction by switching to Sonnet
• Layout generation: 20% of calls → kept on Opus (quality non-negotiable)
• Cache hit rate: ~70% on compliance context → additional 30% savings