Every batch job starts with a number someone remembers. Sometimes it lives in a Slack thread, sometimes in a spreadsheet cell labelled “audio — final”, sometimes in a doc that was accurate three weeks ago. Remembered numbers are not lies; they are photographs of a page that has since moved. For teams pushing tens of thousands of images, audio minutes or weighted tokens through an API, the gap between the remembered figure and the live one is real budget — and it usually surfaces at settlement, after the work is already done.
Why a Remembered Quote Drifts: Provisional Fares and Their Check Dates
Y-PAD publishes its fares inside a beta window, and says so plainly: prices are marked provisional and limited beta, may change before launch, and carry check dates. The list and launch figures in its tables are dated 7 Nov 2026, while the competitor figures beside them were mostly checked 2 Nov 2026 and some 7 Nov 2026. That is not one date for a whole page; it is a date per row — precisely the detail a screenshot throws away. A quote that says “the pricing page said X” without the date is unanswerable when someone asks whether X was a launch rate, a list rate, or a number that has since been revised.
The guarantee that matters for work in flight is narrower and more useful: a submitted batch’s prices are fixed and stored with that batch. So the fare that governs your job is the one on the page at submit time, not the one from the kickoff meeting and not the one that will appear next month. That is an argument for capturing the table at submit rather than trusting recall — and for treating any price older than your current sprint as a rumour until re-checked.
A blog post on the same site shows the drift in miniature. The post at /blog/transcribe-audio-to-srt/, written around 2 Nov 2026, described a $0.0075 minimum per submission and quoted “DeepInfra costs less on a short Whisper job.” The pricing page dated 7 Nov 2026 says there is no minimum charge now and lists DeepInfra slightly above Y-PAD’s per-minute prices, with Y-PAD’s 30-second per-file minimum still applying. Both statements were true once; only one is current. Anyone quoting the blog today would overstate a single short Whisper file by the retired minimum and misname who is cheaper. Five days of edits is all it took to flip the conclusion.
What a Live Table Captures That a Screenshot Cannot
A pasted number loses four things a live table keeps: the launch-versus-list split, the per-file minimums, the unit shapes, and the settlement rule. Each one changes the total.
Launch versus list is a settlement question
At submit, Y-PAD holds the sum of max_units × list price; at settlement you pay for succeeded units at the launch price if you hold a launch seat, otherwise list, and the unused hold returns. Two teams running identical files can therefore settle at different rates. Seats are limited — 40 of them, bought with the $25 pack, applying to the first $100 of usage, one per account and per card, assigned at payment — and if seats run out before payment lands, the pack stays at list prices. Seven models carry two prices, among them Whisper large-v3 ($0.00044 list / $0.0004 launch per audio minute), GLM-OCR ($0.0010 / $0.0005 per page) and Seedance 2.5 ($0.3780 / $0.1850 per video second at 720p). Real-ESRGAN, BiRefNet, Qwen-Image-Edit and Qwen3-TTS carry a single price during the beta.
Minimums and unit shapes are half the quote
Whisper bills by the second with a 30-second minimum per file, so 1,000 ten-second transcripts bill as 30 seconds each — $0.22 list or $0.20 launch — not a third of the one-minute rate. Qwen3-TTS has a 500-character minimum per file, counted after whitespace is folded. DeepSeek V4.1 Flash counts weighted tokens, output plus input at 0.25, with a 100-token minimum per file. Real-ESRGAN prices per image up to 1 MP and pro rata above that to 2 MP, with an alpha channel triggering a second pass at the same cost; BiRefNet goes pro rata up to 4 MP. A per-unit figure quoted without its minimum is not a quote, it is a wish.
Where the comparisons actually land
On the rows where Y-PAD publishes a Replicate figure, Replicate sits above both Y-PAD prices for Real-ESRGAN ($0.002 per image versus $0.0010, checked 7 Nov 2026), BiRefNet (about $0.0017 per run), Qwen3-TTS ($0.02 per 1,000 characters versus $0.0150) and Qwen-Image-Edit (about $0.03 per image versus $0.0120). On video it depends on the settlement rate: Replicate’s $0.2312 per second for Seedance at 720p is above Y-PAD’s launch rate and below its list rate, as is its $0.02 per second for Pruna P-Video against $0.0160 and $0.0210. Whether your batch lands cheaper than replicate therefore depends on the model, the date printed beside the row, and whether you settle at launch or list. No Replicate figure is published for Whisper, DeepSeek V4.1 Flash, GLM-OCR, WAN 2.6 or Qwen3-VL, so no honest quote should imply one.
Building a Quoting Habit That Survives the Next Price Change
The fix is plumbing rather than discipline. Treat a fare the way you treat a pinned dependency: record it with its date and settlement basis, and re-resolve it before a number leaves the building.
Carry four fields with every quote — model id, unit price, check date, and whether it is list or launch. “Real-ESRGAN, $1.00 per 1,000 images, 7 Nov 2026, single beta price” is auditable. “About a cent an image” is folklore by Friday. Store the source page alongside the ticket, not just the number.
Estimate the hold, not only the fare. Because Y-PAD holds max_units × list price at submit, returns unused hold, charges only successful items and bills failed or cancelled items at nothing, your worst-case exposure at any moment is the hold while your actual spend is a settlement event. Teams that budget from succeeded units alone get surprised by their own batch sizes.
Put seat status in the model and keep pack arithmetic visible. Balance never expires; the Starter, Standard and Bulk packs are $10, $25 and $100 as of 7 Nov 2026; the first payment cannot exceed $25, and for seven days after it the day’s charges and open holds together cannot pass $25 — a hard constraint on a large first batch, and a reason to run the small job before the big one.
Price the plumbing too. Cold starts run 10 to 30 seconds, occasionally minutes, with later items in a batch riding the same start; upload links last an hour; a call takes up to 100 files and a batch up to 2,000 items; outputs stay 7 days; there are no webhooks, so you poll; and there is no free storage, so uploads return 402 until the first payment. A quote that omits the polling loop and the re-upload after an expired link is a quote for a different system.
The practical test is a batch ID with an audit trail. 1,000 one-minute transcripts settle at $0.44 list or $0.40 launch as of 7 Nov 2026 — nothing like the retired submission minimum the old blog remembers. In the other direction, 1,000 seconds of Seedance at 720p is $185 at launch and $378 at list, so remembering “launch” without checking seat status understates by more than double. Record what the live pricing page said on the day you submitted, keep it attached to the job, and confirm it again before the next quote goes out.
Pune-raised aerospace coder currently hacking satellites in Toulouse. Rohan blogs on CubeSat firmware, French pastry chemistry, and minimalist meditation routines. He brews single-origin chai for colleagues and photographs jet contrails at sunset.
