Latency budgets become money budgets
Slow SSD-1B retries encourage parallel calls—and parallel charges.
CI & fixtures · SSD-1B
CI should be deterministic. Live SSD-1B calls in tests and Storybook make snapshots flaky and bills noisy—especially for fast mobile and app chrome drafts where visual regression is the whole point.
Slow SSD-1B retries encourage parallel calls—and parallel charges.
Defaulting to the highest SSD-1B tier for every draft wastes premium capacity.
Without a stock layer, every page view path can re-trigger SSD-1B.
Unpredictable SSD-1B lines make early unit economics look worse than they need to.
A PR pipeline regenerates images through SSD-1B for visual tests, email previews, or README builds. Network blips retry jobs; each retry can bill again. Snapshots never stabilize because the model is non-deterministic. Instead, download licensed fixtures once from Epochal, commit or cache them, and assert against fixed bytes. Keep SSD-1B experiments in a manual or nightly job with hard budgets—not on every push. Your teammates get green checks; finance gets a flat image line instead of a surprise inference spike.
Epochal Stock Developer Image API is a single plan at $50 USD/month: full catalog access, 30 requests/minute per account, and unlimited original downloads with one concurrent transfer. There is no per-image fee and no free tier in v1. Use it wherever the picture does not need a unique SSD-1B synthesis—prototypes, fixtures, placeholders—so generative spend stays reserved for finals. Authenticate server-side with a Bearer key only; never put keys in browser JavaScript.
curl -sS "https://epochalstock.com/api/v1/images?tags=workspace,startup&limit=12" \
-H "Authorization: Bearer $EPOCHAL_API_KEY"Keys stay on your server. Never pass ?api_key= or ship secrets to the browser.
Pulled server-side from Epochal tags: server, technology, laptop. Previews are cached; API keys never hit the browser.






Generative outputs are non-deterministic and often metered. CI needs stable fixtures; SSD-1B is for intentional creative runs.
Original downloads are unlimited serially with one concurrent transfer per account. Plan parallel jobs so they do not fight the single-transfer limit.
30 requests/minute per account across all keys. Batch metadata where possible; do not fan out one key per micro-job without backoff.
Yes—gate it behind a manual workflow or quota, not the default PR path. Default fixtures should be stock so SSD-1B spend stays intentional.