Skip to content

cloudflare-cost-guardrails.md

Latest commit

 

History

History
27 lines (19 loc) · 1.73 KB

cloudflare-cost-guardrails.md

File metadata and controls

27 lines (19 loc) · 1.73 KB

Cloudflare Cost Guardrails

This gateway should run without surprise Cloudflare charges. The committed config is intentionally free-first:

  • Workers AI is enabled with WORKERS_AI_ENABLED = "true", but automatic routing keeps it behind non-Cloudflare providers so it is a fallback source of compute.
  • Every Workers AI call path must pass through NEURON_BUDGET, which caps usage at 9,500 Neurons/day, 500 below Cloudflare's 10,000 Neurons/day free allocation.
  • Text and embedding debit estimates use Cloudflare's published per-model token pricing converted back into approximate neurons, plus a 20% buffer. Models without explicit pricing stay on fixed conservative estimates.
  • Workers Logs/observability sampling is disabled in committed config because Workers Logs can create paid overage on Workers Paid plans.
  • Worker CPU is capped at 10ms in committed config, matching the Workers Free per-invocation CPU limit.
  • The unused Cloudflare Rate Limiting binding is not configured; request throttling uses IpRateLimitDO.

Run the local guard before deployment prep:

pnpm audit:cloudflare-costs

pnpm check also runs this audit before typecheck and unit tests.

Do not enable Workers Logs, higher CPU limits, paid-plan-only bindings, or a Workers AI neuron cap above 9,500/day in committed config unless the task explicitly approves paid Cloudflare usage and records the expected monthly ceiling.

Reference points checked on 2026-05-09:

  • Workers AI free allocation: 10,000 Neurons/day.
  • Workers Free request/CPU posture: 100,000 requests/day and 10ms CPU/invocation.
  • Workers Logs paid overage can apply on Workers Paid plans.
  • D1, KV, and SQLite-backed Durable Objects have Free plan quotas that fail closed when exceeded on Free plans.