A price table tells you what a vendor charges for one unit of work. It does not tell you what your batch will cost, how long it will take, or how good the output will be. Those answers live in the billing rules wrapped around the numbers — ceilings, per-file minimums, resolution multipliers, the gap between a list price and a launch price, and the date someone last checked the figure. Treat every line as a guarantee and the gap shows up after settlement. Read the surrounding rules and you get a budget that survives contact with real files.
A Line Item Is a Unit Rate, Not a Promise
The narrow question a rate sheet answers is this: what does one unit cost? For Y-PAD that unit varies by model — per image, per audio minute, per 1,000 characters, per page, per video second, per 1,000 weighted tokens. The second question it answers, if you read the fine print, is what counts as a unit. Both matter, and neither guarantees that a batch of 5,000 mixed files lands on a predictable number.
Take the image models. Real-ESRGAN x4plus is listed at $0.0010 per image up to 1 MP of input, pro rata above that, capped at 2 MP. BiRefNet is $0.0005 per image up to 1 MP, pro rata up to 4 MP. Those two sentences hide two levers: the size of your inputs and the ceiling above which the rate stops being flat. Read without the ceiling, and you would assume a 12 MP photograph costs the same as a thumbnail. It does not. An image with an alpha channel takes a second Real-ESRGAN pass and costs the same again — a detail that doubles the effective rate for exactly the assets a product-photo pipeline tends to hold.
Audio and text carry the same shape of caveat, expressed as per-file minimums rather than pixel ceilings. Whisper large-v3 bills by the second at $0.00044 list / $0.0004 launch per audio minute, with a 30-second minimum per file. Qwen3-TTS has a 500-character minimum per file. DeepSeek V4.1 Flash has a 100-token minimum per file. So 1,000 ten-second transcripts cost $0.22 list / $0.20 launch, not a sixth of the one-minute rate — short files still bill 30 seconds. That is not a hidden fee; it is a unit definition, and it belongs in the estimate.
Dates belong there too. Every figure referenced here is a provisional beta fare published around 7 November 2026, and the vendor states that beta prices may change before launch. A submitted batch keeps the prices stored with it, so the table you read today fixes that batch and nothing after it.
List Price, Launch Price, and the Hold in Between
Many tables at this stage of a product carry two numbers per line, and conflating them is the most common reading error. Y-PAD publishes a list price and a launch price, the latter available to accounts holding one of 40 launch seats. At submit, the platform holds the sum of maximum units multiplied by list price. At settlement, succeeded units are charged at launch price if the account holds a seat, otherwise at list. Unused hold returns, and only successful items are charged — failures and cancellations cost nothing.
That mechanism explains why some lines look dramatically cheaper on a homepage than on the table. Seedance 2.5 image-to-video and text-to-video sit at $0.3780 per video second at 720p on list and $0.1850 on launch. For the same model, the published comparison table lists Replicate at $0.2312 per second at 720p, checked 7 November 2026 — below Y-PAD’s list price and above its launch price. Pruna P-Video has the same geometry: $0.0210 list, $0.0160 launch, and a Replicate figure of $0.02 per second from a third-party listing checked the same day, again falling between the two. Read only the launch column and you overstate the advantage; read only the list column and you understate what a seat actually buys.
The comparison also runs the other way on models where both Y-PAD prices sit under the published alternative. Real-ESRGAN is $0.0010 per image ($1.00 per 1,000) against Replicate at $0.002 ($2.00), and BiRefNet is $0.0005 per image ($0.50 per 1,000) against roughly $0.0017 per run. On those rows the rate is cheaper than replicate at both list and launch — a claim the published tables support, and one that would be wrong to generalise across the whole catalogue, since several models carry no Replicate comparison at all and others sit above competitor pricing. Read the row you actually use, not the headline.
Seats are a commercial condition worth pricing in. They are assigned at payment, one per account and per card, and cover the first $100 of usage at launch price across qualifying models. If seats run out before a payment lands, the pack is kept at list prices. The launch column is conditional; the list column is what applies when the condition fails.
Read the Denominator and the Multipliers Before You Budget
What counts as a unit
Scan the right-hand side of each line before the left. Per minute, per 1,000 characters, per page, per video second and per 1,000 weighted tokens are not interchangeable denominators, and a job that crosses modalities crosses them too. Whisper’s duration is measured from decoded samples rather than the file header, so a badly muxed container will not shrink the bill, and a finished transcription bills even when the transcript returns empty. Qwen3-TTS counts characters after whitespace is folded. GLM-OCR counts pages, with an image input counting as one page and a 100-page ceiling per file. DeepSeek V4.1 Flash runs on weighted tokens — output tokens plus input tokens at a published 0.25 ratio, against a 1,000-token price — which makes the optional max_tokens setting, defaulting to 2,048, a budget lever rather than a formatting preference.
Multipliers, routing and the things a rate cannot promise
Video pricing is where multipliers bite hardest. Seedance 2.5 counts 480p at half, 720p at base, 1080p at 2.5× and 4K at 5×. Pruna P-Video counts 1080p twice. WAN 2.6 counts 1080p at 1.5×. A line that reads “per video second at 720p” is a starting coordinate, not a ceiling. Routing matters as much: Seedance, Pruna and WAN run on Runpod’s hosted endpoints rather than Y-PAD’s own GPUs, and DeepSeek V4.1 Flash runs on Google Cloud Vertex AI, meaning the prompt file leaves the platform’s storage and is sent to Google for each run — a data-handling fact no price column will ever show.
The operational envelope is the other half of the cost. Cold starts of 10 to 30 seconds are normal and occasionally stretch to minutes, with later items in a batch riding on the same start. There are no webhooks, so a 2,000-item batch means polling. Outputs stay seven days, upload links last one hour, and a single call takes up to 100 files. None of that appears in a per-unit column, and all of it shapes how much engineering time a “cheap” line item really consumes before the first download lands.
Cairo-born, Barcelona-based urban planner. Amina explains smart-city sensors, reviews Spanish graphic novels, and shares Middle-Eastern vegan recipes. She paints Arabic calligraphy murals on weekends and has cycled the entire Catalan coast.
