Almost every bulk AI pipeline is really two pipelines stacked on top of each other. The first is perception: turning pixels, waveforms and page scans into plain text. The second is reasoning: taking that text and deciding what it means. Teams frequently buy both halves as one product, then discover the second half is doing three surprisingly different jobs — attaching a label, pulling fields into a schema, and compressing a long document into something a human will actually read.

All three jobs share the same request shape. Text goes in; a constrained string comes out. That is exactly what a text-generation endpoint offers, which is why classification, extraction and summarisation so often belong on the same model call rather than three separate services. The interesting question is not whether they fit — it is where they fit, and where the fit breaks.

When the Job Is Just Text In, Constrained Text Out

Classification, extraction and summarisation look different on a product spec, but once the input is tokenised they converge. Classification returns a label drawn from a fixed set. Extraction returns an object whose keys you already know. Summarisation returns prose under a length budget. In each case the quality of the result depends on the same three things: a clear instruction, a bounded output, and a representative sample to test against.

That convergence is why a single text-generation endpoint covers so much ground in bulk jobs. Y-PAD’s model page for DeepSeek V4.1 Flash describes the model with four verbs — classify, extract, translate, summarise — and it accepts either plain text or a JSON messages array with up to 64 system, user and assistant entries. Output can be plain text or JSON, and max_tokens is settable from 1 to 8,192 with a default of 2,048. The JSON mode is the part that makes extraction practical at scale: your parser, not a human, reads the response, so a malformed object fails loudly instead of silently landing in a spreadsheet.

Cost modelling on a text endpoint is a different discipline from per-image or per-minute billing. The DeepSeek api on Y-PAD is sold per 1,000 weighted tokens, where output tokens count fully and input tokens count at 0.25 — the published price ratio. That ratio matters: long prompts are cheap, long answers are not. A summarisation job that asks for 600 words of prose will be dominated by output cost, while a classification job over a 5,000-word ticket will be dominated by the input ratio and still land well below a cent per file at the published rate of $0.0015 list or $0.0009 launch per 1,000 weighted tokens.

There is one floor worth knowing: a 100-token minimum per file. Short items are not free. Ten thousand one-line classifications still bill as if each carried 100 tokens, so the minimum, not the prompt, sets your cost per unit. Y-PAD also applies batch pricing to every request — one file costs the same per unit as a thousand — which makes it cheap to run a single file through the playground and the API before committing a whole corpus.

Where the Text Endpoint Should Hand Off to a Specialist

A text-generation endpoint cannot see. If your invoice arrives as a scan, or your meeting as an audio file, the language model is the wrong first hop. The pipeline should chain perception into reasoning rather than asking one model to do both badly.

For documents, Y-PAD’s GLM-OCR converts a PDF or image page into Markdown at $0.0010 list or $0.0005 launch per page, one page per image and up to 100 pages per file. For images that need structured answers rather than transcribed text, Qwen3-VL 8B returns JSON or text per image at $0.0020 list or $0.0010 launch, answers capped at 2,048 tokens. For audio, Whisper large-v3 produces json, text, srt or vtt at $0.00044 list or $0.0004 launch per audio minute, billed by the second. In each case the specialist extracts the literal content, and the language model does the judgment work on top.

The same hand-off logic applies inside text. If a label can be derived by a regex, an allowlist or a numeric threshold, a model call is a rounding error on a solved problem. Extract the deterministic fields first and reserve the endpoint for the genuinely ambiguous residue — the complaint that could be billing or could be churn, the clause that changes the meaning of the contract.

There are governance edges too. Y-PAD documents that DeepSeek V4.1 Flash runs on Google Cloud Vertex AI rather than Y-PAD’s own GPUs, so each run sends the prompt file out of Y-PAD storage to Google. Routing a scanned HR document through a perception model and then a text model is a data-flow decision, not just a cost decision. Check who sees what at each hop before you batch a sensitive corpus. And confirm live pricing before committing, because the fares above are provisional beta prices as of the dates Y-PAD publishes.

Designing the Batch: Holds, Polling and Structured Output

Once the job type is settled, the batch mechanics decide whether the run is smooth or painful. There are no webhooks on Y-PAD; you poll. Upload links last one hour, a call accepts up to 100 files, and a batch can hold up to 2,000 items. Outputs remain downloadable for seven days, which is long enough to reconcile results and short enough that you should not treat it as storage. Cold start runs 10 to 30 seconds and occasionally minutes, so a single large batch is usually friendlier than dozens of small ones.

Budgeting is where structured output earns its keep. At submit, Y-PAD holds the sum of max_units multiplied by 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 to your balance. That means max_tokens is not just a quality dial, it is a hold ceiling. Setting it to 256 for a classification job instead of leaving the 2,048 default frees a large chunk of prepaid balance for other work in flight.

Launch seats are the other lever. Y-PAD publishes 40 seats, claimed by buying the $25 pack, covering the first $100 of usage at launch price, one seat per account and per card. DeepSeek V4.1 Flash is among the models they apply to, at any request size, on succeeded units only. Its own comparison table places DeepSeek’s official peak weighted rate at $1.20 per million tokens — below Y-PAD’s list figure and above its launch figure — so a seat changes the arithmetic on a large extraction run but not on a small one. Only successful items are charged, so failed and cancelled items cost nothing.

Before scaling, prove the schema on a hundred files in the playground, watch the failure modes rather than the successes, and keep the prompt’s output contract narrow enough that a JSON parse error is the loudest thing that can go wrong.