Most bulk text jobs do not overspend because the model was wrong. They overspend because the estimate used the wrong unit. A prompt that looks short can carry thousands of input tokens, while a one-line answer can sit behind a generous max_tokens ceiling. When a provider prices text generation per weighted token, the fare blends what you send with what the model writes back — and that blend is what you must forecast before you submit. With batches holding up to 2,000 items, small per-file errors compound quickly.

What a Weighted-Token Fare Actually Counts

DeepSeek V4.1 Flash is quoted per 1,000 weighted tokens: $0.0015 at list and $0.0009 at launch as of 7 November 2026, both marked as provisional beta fares. The rule behind the weighting is straightforward — output tokens count in full, input tokens count at 0.25. A file that sends 2,000 input tokens and produces 400 output tokens therefore bills as 400 + (2,000 × 0.25) = 900 weighted tokens. A 100-token minimum applies per file, so tiny prompts still pay a floor even when the raw arithmetic lands at 30 or 60.

The ratio tells you where to optimise. Trimming a long instruction block saves a quarter of a token for every token removed; shortening the answer saves a full token for every token removed. If a batch is already producing concise JSON, most of the remaining billable weight usually sits in the prompt. That is why the same model can be cheap for extraction and expensive for summarisation, even when both jobs look similar on a spreadsheet.

Input arrives as text/plain or as an application/json messages array carrying at most 64 system, user and assistant texts. Output can be plain text or JSON, and max_tokens runs from 1 to 8,192 with a default of 2,048. The model runs on Google Cloud Vertex AI rather than Y-PAD’s own GPUs, so the prompt file leaves Y-PAD storage and is sent to Google for each run — worth noting if your budget review also touches data-handling review. The model page for DeepSeek v4 api documents these input shapes and limits.

A one-thousand-file example makes the arithmetic concrete. At 2,000 input tokens and 400 output tokens per file, each file weighs 900 tokens, so the batch totals 900,000 weighted tokens, or 900 billing units. At list that is $1.35; at launch it is $0.81. Neither figure includes a per-request minimum, because Y-PAD applies batch pricing to every request — one file costs the same per unit as a thousand. The per-file minimum is a different thing, and it still bites on short jobs.

The Hold, the Settlement, and Why Your Balance Looks Smaller Than the Bill

At submit, Y-PAD holds the sum of max_units × list price across the batch. For a token-priced model, the maximum billable units follow the ceiling you set, so a default max_tokens of 2,048 reserves far more balance than a tight ceiling of 300. The hold is not the bill. At settlement you pay for succeeded units only, at launch price if you hold a launch seat and otherwise at list, and unused hold returns to your balance. The prices used are stored with the batch, which means a later fare change does not rewrite a settled job.

This design rewards a two-step workflow. Submit a pilot batch with realistic ceilings, read the settled cost, then reset max_tokens per job type before the full run. Classification and routing jobs rarely need more than a few hundred output tokens; long-form summaries need headroom. One ceiling for everything either over-reserves or truncates.

Prepaid packs shape the cash side of the budget. As of 7 November 2026 the packs are Starter at $10, Standard at $25 and Bulk at $100, and the first payment cannot exceed $25. For seven days after that first payment, the day’s charges plus open holds cannot pass $25. That cap is easy to hit with a large batch submitted at list-price holds, so sequence exploratory runs before high-volume ones rather than launching everything in a single call.

Two smaller rules matter when you model failure rates. Failed and cancelled items cost nothing, so a batch that partially succeeds bills partially — budget for the successes rather than the submission count. And there is no minimum charge at the request level, so splitting work across many small requests does not add a per-request fee. Only the per-file token minimum, the 100-token floor, survives at the item level.

Forecasting a Real Batch: Tokens, Minimums and the Launch Seat

Three patterns cover most text batches. First, product descriptions at 200 input and 150 output tokens weigh 200 tokens per file; 1,000 files equal 200,000 weighted tokens, or 200 units, costing $0.30 at list and $0.18 at launch. Second, short classification at 80 input and 20 output weighs only 40 tokens, so the 100-token minimum applies; 1,000 files bill as 100,000 weighted tokens, $0.15 at list and $0.09 at launch. The floor more than doubles the effective rate on that job, which is an argument for consolidating records into one file using the 64-text messages array instead of sending one record per file.

Third, long summarisation at 8,000 input and 800 output tokens weighs 800 + 2,000 = 2,800 tokens per file. A thousand files reach 2.8 million weighted tokens, or 2,800 units, for $4.20 at list and $2.52 at launch. The input-heavy shape is where the 0.25 ratio earns its keep: without weighting, the same job would bill 8.8 million tokens instead.

The launch seat changes the multiplier, not the arithmetic. Y-PAD publishes 40 seats, assigned at payment when you buy the $25 pack, covering the first $100 of usage at launch price, one seat per account and per card. If seats run out before payment lands, the pack is kept at list prices. Treat launch pricing as a temporary planning assumption, not a permanent line in your model. Y-PAD’s own comparison table puts DeepSeek’s official peak weighted rate at $1.20 per million tokens, which sits below Y-PAD’s list rate of $1.50 and above the $0.90 launch rate — the homepage “25% under DeepSeek” claim refers to launch pricing only.

Finally, plan the mechanics around the budget. Upload links last one hour, calls accept up to 100 files, batches hold up to 2,000 items, there are no webhooks, and cold starts of 10 to 30 seconds are normal. Outputs stay available for seven days. Check the live pricing page before committing, because these fares are provisional beta figures with published check dates. Then log weighted tokens per file from the settled batch; the next estimate comes from measured output rather than a ceiling.