A single key in front of several models reads as convenience: one credential, one balance, one place to download results. In practice it is a bundle of separate decisions wearing a trench coat. Before the key lands in a worker queue, decide whether “one key” is meant to deliver one credential, one invoice, one request shape, one data boundary, or all four. Those promises do not all travel together, and whichever one you skip is the one that surfaces at an inconvenient hour.
Credential scope comes first. When a key is displayed once at creation and never again — the shape Y-PAD uses, with keys prefixed op_live_ — the entire rotation story has to exist before the first upload. One key per environment, never one key shared across a team, stored in a secret manager rather than a notebook cell, with a documented revocation path that does not depend on someone remembering a dashboard URL.
Licence scope comes second. A unified key does not unify licences. The same catalogue holds Real-ESRGAN under BSD-3-Clause, BiRefNet under MIT, Whisper large-v3 under MIT (OpenAI) with an MIT-licensed faster-whisper conversion, Qwen3-TTS under Apache-2.0, and DeepSeek V4.1 Flash, whose licence sits with Vertex AI Model Garden under publisher deepseek-ai. If outputs feed something you redistribute, that per-model list is the document that matters — not the key settings page.
Data boundary comes third. “One API” is not “one data path.” The video models — Seedance 2.5 image-to-video and text-to-video, Pruna P-Video and WAN 2.6 — run on Runpod’s hosted endpoints rather than Y-PAD’s own GPUs, and DeepSeek V4.1 Flash runs on Google Cloud Vertex AI, which means the prompt file leaves Y-PAD storage and is sent to Google for each run. Decide per job class which boundary is acceptable before routing a thousand files through it.
Schema drift comes fourth. JSON shaped like Replicate’s keeps migration cheap, but cheap migration is not stable naming. Pin model ids such as real-esrgan, whisper-large-v3 or deepseek-v41-flash, and treat marketing names — “DeepSeek api” in one ticket, “DeepSeek v4 api” in another — as aliases your code resolves, never as identifiers your database stores.
Decide Your Cost Unit, Your Ceiling and Your Reconcile Path
Estimators written before anyone reads the billing page are the most common source of blown budgets in bulk work, because “one unit” means something different on nearly every model. One catalogue can contain per-image pricing (Real-ESRGAN at $0.0010 per image up to 1 MP, pro rata to 2 MP; BiRefNet at $0.0005 up to 1 MP, pro rata to 4 MP), per-edited-image pricing (Qwen-Image-Edit at $0.0120), per-page pricing (GLM-OCR at $0.0010 list or $0.0005 launch per page), per-character pricing (Qwen3-TTS at $0.0150 per 1,000 characters), per-audio-minute pricing (Whisper large-v3 at $0.00044 list or $0.0004 launch, billed by the second), per-1,000-weighted-token pricing (DeepSeek V4.1 Flash at $0.0015 list or $0.0009 launch, output tokens plus input at the published 0.25 ratio), and per-video-second pricing with resolution multipliers. Those are provisional beta fares as of 7 Nov 2026, so confirm them against the live pricing page before a forecast leaves your laptop.
Then separate per-request minimums from per-file minimums, because they behave differently. Every request gets batch pricing with no minimum charge — one file costs the same per unit as a thousand — but per-file floors still bite: 30 seconds on Whisper, 500 characters on Qwen3-TTS counted after whitespace is folded, 100 tokens on DeepSeek V4.1 Flash. Ten thousand six-second voice notes are not 60,000 seconds of work in the estimator’s eyes; they are 300,000 billed seconds.
Decide your ceiling next. At submit, Y-PAD holds the sum of max units times list price, settles succeeded units at launch price if the account holds a launch seat and at list otherwise, and returns the unused hold, with the prices used stored alongside the batch. That mechanic makes ceilings a budgeting decision, not just a safety valve: the maximum you set is the amount held while a 2,000-item batch is in flight. Prepaid packs never expire, the first payment cannot exceed $25, and for seven days after it, charges plus open holds are capped at $25. Launch seats are limited to 40, assigned at payment with one per account and per card, and cover the first $100 of usage at launch price. The docs behind Ai-api spell out request and settlement shapes, which is a better place to model the hold math than a spreadsheet reconstructed from invoices.
Decide What Happens to Work in Flight: Timing, Failure and Retention
Bulk operators underestimate queue mechanics far more often than model quality. Expect a cold start of roughly 10 to 30 seconds, occasionally minutes, and remember that later items in a batch ride on the same start — so a hundred-file call is not a hundred cold starts, but it is also not a stream of results arriving one by one. There are no webhooks; you poll. Design the worker around that now or rebuild it after the first overnight run.
Decide your retention story too, since nothing here is a filing cabinet. Outputs stay available for seven days, upload links last one hour, calls carry up to 100 files and batches up to 2,000 items, and there is no free storage — uploads return 402 until the first payment lands. A pipeline that assumes it can re-download last month’s upscales cannot.
Failure rules are where unified billing earns its keep: only successful items are charged, and failed or cancelled items cost nothing. But “successful” has model-specific edges worth deciding on early. An image with an alpha channel takes a second pass through Real-ESRGAN and costs the same as the first, so a thousand transparent PNGs is two thousand billed images. Whisper bills a finished transcription even when the transcript comes back empty, and measures duration from decoded samples rather than the file header — which quietly moves estimates for oddly encoded audio. Qwen3-TTS offers preset voices only, with no cloning and no reference audio, and a 20,000-character ceiling per file after whitespace folding, so a “clone the narrator” requirement belongs in model selection, not retry logic. DeepSeek V4.1 Flash accepts at most 64 system, user and assistant texts and caps max_tokens between 1 and 8,192, defaulting to 2,048.
Finally, decide how you will reason about vendor comparisons without fooling yourself. Y-PAD publishes its own Replicate figures with check dates, and they are not uniform: on Real-ESRGAN its table shows $1.00 per 1,000 images against Replicate’s $2.00 (checked 7 Nov 2026), while on Seedance 2.5 and Pruna P-Video the Replicate per-second figures sit between Y-PAD’s launch and list fares. Budget at list unless a launch seat is confirmed, and read the dates rather than the percentage claims.
