DeepSeek API TWD Cost: Sept 2026 Prices and Usage Estimates

Price query date: 2026-09-17 (Taipei time). This article is a customer-price estimate based on BazaarLink's public model catalog. It is not DeepSeek's official direct-purchase price, nor an upstream cost table. The TWD conversion uses a fixed US$1 = NT$31.5 for comparison; it is not a live exchange rate or the actual checkout rate.

How much DeepSeek API costs depends first on which full model ID you use, and then on the tokens sent with each request. Based on BazaarLink's current V4 Flash prices of US$0.20 input and US$0.40 output per million tokens, a customer-service bot handling 500 conversations per day with three turns each costs about NT$170 per month under the assumptions below. This is an estimate, not an actual customer-service bill, and it is not the case that every customer-service operation can be kept within NT$100.

If you want to compare DeepSeek's official pricing with Taiwan platforms, put currency, model version, caching, service fees, and document requirements in the same table. Seeing only the word "DeepSeek" does not mean two products are identical.

BazaarLink Model Cost Comparison (USD and NT$)

The table below matches each item against the pricing.prompt and pricing.completion fields in the BazaarLink public API model catalog. The source fields are US dollars per token; the table multiplies them by one million. The catalog can also be viewed on the Models page. Models with similar names or dates are separate entries, so their prices cannot be mixed.

This does not rank the models by capability. For limited-budget summaries, customer-service drafts, and batch processing, test Flash first. For multi-step reasoning, compare Pro or R1 using the same set of questions. Output that is cheap but has to be redone by hand can end up costing more.

Three Recalculable Monthly Cost Examples

All calculations assume 30 days. No caching discount is assumed, and no usage rebate is deducted. The following are scenario examples with stated usage; these three production workloads were not actually run.

Monthly cost US$＝monthly input tokens ÷ 1,000,000 × input price
         ＋monthly output tokens ÷ 1,000,000 × output price
TWD comparison value＝monthly cost US$ × 31.5

Customer-Service Bot: V4 Flash, 500 Conversations per Day

Assume each conversation averages three turns, with an average of 300 input tokens actually sent by the API each turn, and an average of 150 billed output tokens.

300 tokens is not the length of "the sentence the customer just typed." The system prompt, knowledge base excerpts, and message history can all be part of the input. If the actual average input per turn is 3,000 tokens, with everything else unchanged, the monthly cost becomes US$29.70, about NT$935.55. Pulling one week of real usage data is more useful than guessing at conversation length.

Nor should you borrow a "customer service costs about X" figure from another article to compare multiples. Model IDs, output lengths, or price dates differ between articles, so the resulting difference may not come from the model itself. If you need to actually reduce costs, see customer-service model routing and cost control.

Document Batch: V4 Flash, 100 Documents per Day

Assume each document uses 6,500 API input tokens and an 800-token summary output. There is no fixed conversion ratio between word count and tokens, so token usage is specified directly here.

• Monthly input: 100 × 30 × 6,500＝19.5 million tokens; cost US$3.90.
• Monthly output: 100 × 30 × 800＝2.4 million tokens; cost US$0.96.
• Total US$4.86, about NT$153.09 per month.

The free quota can be used to verify your setup first and spot-check output quality, but do not assume that a batch job of 100 documents per day runs entirely free. Set the formal budget using the actual limits and paid terms.

Reasoning Task: R1, 200 Questions per Day

Assume each question uses 2,000 input tokens and 4,000 billed output tokens. For reasoning models, estimate from the billed usage the API reports, not only from the length of the visible answer.

• Monthly requests: 200 × 30＝6,000.
• Monthly input: 12 million tokens × US$0.70＝US$8.40.
• Monthly output: 24 million tokens × US$2.50＝US$60.00.
• Total US$68.40, about NT$2,154.60 per month.

Output cost accounts for most of this example. Shortening unnecessary answers and separating the questions that truly need reasoning controls the budget better than focusing only on the input price.

Free Tier: Check the Full ID and Quota First. Seeing free Does Not Mean Zero Cost

As of the query date, the zero-price DeepSeek entry in the public catalog is deepseek/deepseek-v4-flash-0731free:free. There is also an entry containing 0731free that lists a normal paid price. A name containing free is not proof of free use. You must check the full ID, the catalog, and the free-tier rules.

To verify your API setup first, you can use auto:free. However, it selects whichever free model is available at the moment and does not guarantee DeepSeek every time. For live candidates and baseline limits, see the free models page. At the time of this query there were two models, with a baseline of 10 requests per minute and 50 per day, plus an account-tier multiplier.

Actual eligibility and multipliers are determined by your account balance, subscription status, and the platform's current thresholds. After the free quota is exceeded, requests that meet the paid fallback conditions may be executed at the normal rate; requests that do not meet them receive a rate limit. Do not treat a free test on a topped-up account as something that will never be charged, and long-running batch jobs should not rely solely on the free quota.

Usage Rebate: Which Spending Counts and When It Is Credited

The DeepSeek model IDs covered by the current usage rebate description are deepseek-v4-flash, deepseek-v4-flash-0731, deepseek-v4-pro, and deepseek-v4-pro-0813. R1, V3.2, and V4.1 Flash are not on the list found for this query.

The fixed rebate at each threshold accumulates, and the portion above US$500 is calculated at 20%:

For example, spending US$49 keeps the cumulative rebate at US$4; spending US$100 brings it to a cumulative US$20; spending US$500 brings it to a cumulative US$100. "20%" does not mean every purchase is immediately discounted by 20%.

None of the three estimates above should have the rebate deducted first. The customer-service US$5.40 and document US$4.86 both fail to reach the first threshold, and the R1 monthly cost of US$68.40 is for a model that is not on the plan list.

Crediting timing also needs to be understood separately. Milestone rebates settle daily and are credited about the next day after the milestone is reached. The proportional rebate after the final tier is credited on the 1st of the following month, based on the UTC calendar month. Progress is recalculated each month. Only personal and prepaid organizations are eligible; monthly-billed organizations are not. Free usage, BYOK usage, and usage offset by subscription points do not count. If an organization already has a valid contract discount on the same model group, it cannot be combined with the usage rebate. Rebates are balance credits: they are not refunded, do not generate a separate invoice, and do not expire. When rules change, confirm against the current plan, your account, and your organization contract.

BazaarLink vs. DeepSeek Official Direct Purchase: How to Compare Meaningfully

For DeepSeek's official current models, currencies, and caching and time-of-day pricing, check the official API pricing directly. Do not cross-calculate official direct prices with the customer prices in the table above, and do not treat similarly named versions as the same price entry.

If you only use DeepSeek and already have a suitable payment method, first compare the full official cost. If you need multi-model management and Taiwan transaction documents, then evaluate whether BazaarLink is worth it. Whether overseas receipts can be expensed, and how input tax on a unified invoice is handled, are separate questions; confirm them with your company's accountant before purchasing. For the process, see the AI API expense reporting guide for Taiwan.

Actual payment may also be affected by top-up service fees, taxes, and currency conversion, so the TWD estimates here do not replace the checkout amount. Fill the formula with your own input and output usage first, then decide which model to buy. For the latest prices, see the model catalog.

TWD billing · Taiwan invoices · leading AI models · OpenAI-compatible API