How I built pairwise AI model compare pages with Claude Haiku and a budget cap

Short summary

A developer built pairwise comparison pages for 200+ AI models using Claude Haiku while managing costs. The solution groups models by pipeline type, takes the top 4 per group, and caps comparisons at 50, turning a potential 20K pairs into ~48 viable comparisons. The implementation uses prompt caching for cost reduction, deterministic slugs for idempotency, and defensive JSON parsing with fallbacks.

• •Group models by pipeline type; generate comparisons only for top performers within each group
• •Use prompt caching on the system prompt (JSON schema) to reduce token cost across multiple calls
• •Make the entire ETL idempotent: deterministic slug generation + explicit existence checks + fallback schema handling

Generated with AI, which can make mistakes.