Representative set
Cover the categories and nasty prompts customers will send.
Benchmarks you can reproduce.
First-party image API benchmarks
Methods for measuring image API cost, latency, and reliability without disappearing failed runs, warm-ups, or discard rate from the chart.
How to use these guides
A useful benchmark records every condition that could move the number.
One lucky success is not evidence. Use a representative prompt set, hold dimensions constant, separate warm-ups, keep failures, and report median, p95, sample size, model, date, and cost source.
Complete data
Keep rejections, timeouts, retries, and discards in the math.
Publish context
State size, date, model, region, sample size, cost source.
From guide to request