Cost Optimization

Cost Optimization

Review Checklist

• AI calls use cheapest viable model (Haiku for classification, Sonnet for most tasks, Opus only for complex judging)
• System prompts identical across calls (enables Anthropic cache — 90% savings)
• Database queries SELECT only needed columns (not select('*'))
• No polling when webhooks/Realtime could work
• Cost tracked per feature in api_costs table
• Batch API used for non-urgent AI work (50% cheaper)

Cost Breakdown by Service

Vercel

Optimizations:

• runtime = 'edge' for simple routes (cheaper than Node.js runtime)
• next/image optimization reduces bandwidth significantly
• Incremental builds with Turborepo — only rebuild changed packages

Supabase

Optimizations:

• SELECT only needed columns (reduces egress)
• Paginate queries (never unbounded)
• Compress transcripts (gzip → 80% size reduction)
• Expire idle Realtime channels
• Use materialized views for leaderboards (compute once, read many)

Anthropic API (Biggest Variable Cost)

Optimizations:

• Prompt caching: Identical system prompts are cached by Anthropic — 90% savings on input tokens. Keep system prompts consistent across calls.
• Batch API: 50% cheaper for non-urgent work. Arena judge scoring is a perfect candidate — results don't need to be instant.
• Model tiering: Don't use Sonnet for what Haiku can do. Don't use Opus for what Sonnet can do.
• Cost tracking:

// Log every AI call with cost estimate
async function trackedAICall(model: string, fn: () => Promise<APIResponse>) {
  const result = await fn()
  const cost = estimateCost(model, result.usage.input_tokens, result.usage.output_tokens)
  
  await supabase.from('api_costs').insert({
    service: 'anthropic',
    model,
    feature: 'arena_judge', // tag by feature
    input_tokens: result.usage.input_tokens,
    output_tokens: result.usage.output_tokens,
    estimated_cost_usd: cost,
  })
  
  return result
}

// Daily rollup + budget alert
// pg_cron: if daily spend > $50, alert Nick

Cost-Per-Feature Tracking

Budget rule: Set monthly spending caps per service. Alert at 80%. Hard limit at 100%.

Sources

• Vercel pricing documentation
• Supabase pricing documentation
• Anthropic API pricing (2026)
• Resend pricing

Changelog

• 2026-03-21: Initial skill — cost optimization