cost-benchmark

Reproducible GPT Image 2, Flux, Gemini and Nano Banana image API benchmarks with raw outputs, prompt fixtures, latency and cost data.

Event-driven LLM request pipeline with a Postgres queue, idempotent ingress, retry/DLQ/replay — and the paired benchmark showing Haiku 4.5 cost 2-7x more per task than Sonnet 5 despite half the per-token price, which is why there is no cost router in src/.