Queue four hundred research interviews and the first thing that stops the job is rarely the model. It is the reserve placed against the account balance at submit time. For teams running bulk transcription of oral-history archives, lecture capture libraries or podcast back catalogues, that reserve is what determines whether the batch runs tonight or after the next top-up. Treating it as a guess is the mistake; on a published pricing table it is arithmetic you can do before you upload.
The Hold Is Arithmetic, Not a Guess
At submit, Y-PAD holds the sum of each item’s max_units multiplied by list price. That number is a ceiling, not a quote. At settlement, succeeded units are charged at launch price if the account holds a launch seat and at list otherwise, and whatever remains of the hold goes back to the balance. Only successful items are billed: failed and cancelled items cost nothing, and prepaid balances never expire. The prices used are stored with the batch, so a later price change does not rewrite what a submitted job was priced at.
Two details routinely surprise people sizing audio batches. First, duration is measured from decoded samples, not from the file header. A recorder that writes a confident header over a truncated WAV is still billed on the audio that actually decodes. Estimate hold from a metadata duration field and your ceiling drifts, sometimes badly. Probe the real stream before you submit and build minutes from what the decoder reports.
Second, per-file floors and caps move the ceiling more than the headline rate does. Whisper large-v3 bills by the second at list $0.00044 and launch $0.0004 per audio minute as of 7 November 2026, with a 30-second minimum per file, a 240-minute maximum per file and a 500 MB file limit. A finished transcription bills even when the transcript comes back empty, because silence is still duration. A folder of 12-second voice notes and a folder of hour-long interviews can share identical total playtime and still produce very different bills.
Because these are provisional beta fares with a check date, confirm current figures on the live pricing page before sizing a large project. The useful part for planning is that the formula is published, not that any single number is permanent.
Sizing Batches Around Caps, Packs and the First-Week Rule
Most budgeting problems appear in the first week rather than on the thousandth file. The first payment cannot exceed $25, and for seven days after it the day’s charges plus open holds cannot pass $25 together. That is a hard concurrency constraint: a single submission whose hold alone approaches the ceiling leaves no room for anything else to run while it settles. Plan the opening days as a sequence of smaller batches that drain as they complete, rather than one monolithic job sitting open.
Pack structure sets the floor. Starter is $10, Standard $25 and Bulk $100, and the $25 pack is the one that claims a launch seat while seats remain — 40 seats in total, one per account and per card, assigned at payment, covering the first $100 of usage at launch price on qualifying models including Whisper large-v3, any request size, on succeeded units. If the seats run out before a payment lands, the pack is kept and charged at list. For a long transcription project that seat is the gap between $0.44 and $0.40 per audio minute — trivial per file, obvious across an archive.
Structural limits determine how you slice a folder: up to 100 files per call, up to 2,000 items per batch, outputs retained for 7 days, upload links valid for one hour. There are no webhooks, so you poll for status. Cold starts of 10 to 30 seconds, occasionally minutes, apply per batch, and later items ride the same start — so co-locating related work in one submission beats trickling files in one at a time.
Splitting long recordings is the cleanest lever you have. The 240-minute per-file ceiling is both a hard limit and a budgeting tool: a file sitting near the cap makes that item’s hold the worst case you could ever be charged. Cutting an interview at chapter boundaries or silence points keeps max_units close to real duration, which keeps the hold close to the settled cost. The split point becomes your choice instead of something discovered on the invoice.
Worked Numbers for Interview Archives and Lecture Libraries
Y-PAD’s published examples make quick sanity checks easy. As of 7 November 2026, one thousand one-minute transcripts come to $0.44 at list or $0.40 with a launch seat; one thousand ten-second transcripts come to $0.22 at list or $0.20 at launch, because those short files still bill the 30-second minimum. Scale that up: 300 podcast episodes averaging 40 minutes is roughly 12,000 audio minutes, about $5.28 at list or $4.80 at launch, spread across three calls to respect the 100-files-per-call limit. The hold is computed at list and partly returns on settlement. Accepted containers and output formats — json, text, srt, vtt — are documented on the whisper api page.
Per-file minimums are where bulk jobs quietly drift. The same pattern recurs across the model list: Qwen3-TTS enforces a 500-character minimum per file, DeepSeek V4.1 Flash a 100-token minimum per file on weighted tokens, Whisper its 30-second floor. Batch thousands of tiny items and those floors dominate, so an accurate hold looks wrong against a naive per-second or per-character estimate. Either pool short items into single files where the workflow permits, or price the batch off item count rather than total content.
Image and document jobs follow identical logic with different units. Real-ESRGAN at $0.0010 per image up to 1 MP and BiRefNet at $0.0005 per image up to 1 MP, both pro rata above that, mean a hold depends on pixel area and on how many assets carry alpha channels, since an alpha image takes a second pass at the same price. GLM-OCR bills per page up to 100 pages per file, so a 90-page PDF is one item, not ninety — a distinction that changes a hold by two orders of magnitude if you get it backwards.
Operationally, this collapses into one function in your submission code: probe or decode each file, compute max_units per item, multiply by list price, sum, then compare against what the balance can actually carry under the current constraints. Run a single file through the playground first to confirm the shape of the response, then let the script do the reserving. Estimate before submit, reconcile after settlement, and the hold stops being a number anyone has to guess.
Belgrade pianist now anchored in Vienna’s coffee-house culture. Tatiana toggles between long-form essays on classical music theory, AI-generated art critiques, and backpacker budget guides. She memorizes train timetables for fun and brews Turkish coffee in a copper cezve.