Pricing Example: DeepSeek V4 Flash

Applicable roles: Admin user, Member user Last updated: 2026-08-06

This article uses deepseek-v4-flash to demonstrate how fees are calculated. The example numbers are for illustration only and do not represent current official prices. For actual unit prices, discounts, and billing dimensions, refer to the model plaza and "Usage Records > Fee Details."

1. Example Call

Assume a single call contains:

• Model: deepseek-v4-flash
• Group discount: 0.8
• Input: 10,000 tokens (of which 6,000 tokens hit the prompt cache)
• Output: 2,000 tokens
• Cache write (5-minute): 4,000 tokens (the portion written to cache for the first time)
• Cache read: 6,000 tokens (the portion that hit the cache)
• Input standard price: 2 credits / million tokens
• Output standard price: 8 credits / million tokens
• Cache write standard price: 3 credits / million tokens
• Cache read standard price: 0.2 credits / million tokens

The unit prices and discount above are demonstration values. Do not use them for procurement, quotation, or financial settlement.

2. Calculation Formula

Per-item fee = Token count / 1,000,000 x Standard unit price x Group discount
Total fee = Input fee + Output fee + Cache write fee + Cache read fee + Other enabled dimension fees

Calculated with the example data:

Input fee       = 10,000 / 1,000,000 x 2   x 0.8 = 0.016   credits
Output fee      =  2,000 / 1,000,000 x 8   x 0.8 = 0.0128  credits
Cache write fee =  4,000 / 1,000,000 x 3   x 0.8 = 0.0096  credits
Cache read fee  =  6,000 / 1,000,000 x 0.2 x 0.8 = 0.00096 credits
Total fee       = 0.016 + 0.0128 + 0.0096 + 0.00096 = 0.03936 credits

Notes:

• Input is billed on the full token count (including the portion that hit the cache). Hitting the cache does not reduce the input fee.
• Cache write is charged only on the first write. Subsequent requests that reuse the cache do not incur a write fee. Cache writes have two TTL tiers -- "5-minute" and "1-hour" -- with potentially different unit prices. Refer to fee details for specifics.
• Cache read covers the portion that hit an existing cache. Its unit price is much lower than regular input, making it the primary cost-control lever for long-context scenarios.
• Token counts and per-item fees in usage records may be subject to display-precision rounding. Manually summing individual items may produce a tiny rounding difference from the actual charge. The "actual total charge" in fee details is authoritative.

3. How Discounts Apply

Example fees without discount:

Input       = 10,000 / 1,000,000 x 2   = 0.02
Output      =  2,000 / 1,000,000 x 8   = 0.016
Cache write =  4,000 / 1,000,000 x 3   = 0.012
Cache read  =  6,000 / 1,000,000 x 0.2 = 0.0012
Standard total = 0.02 + 0.016 + 0.012 + 0.0012 = 0.0492 credits

After applying the 0.8 group discount, the actual payment is 0.03936 credits -- a 20% saving compared to the standard price.

Group discounts may be adjusted, but historical records retain the price and discount in effect at the time of billing for audit and reconciliation purposes.

4. Possible Billing Dimensions

Image, audio, and video inputs and outputs are each priced independently.

Not all models support caching or multimodal input. Do not assume that deepseek-v4-flash and kimi-k3 share the same billing dimensions. Refer to the console for confirmation.

5. How to Control Costs

1. Create a separate API Key for each business unit or individual.
2. Configure balance limits, quota limits, rate limits, and concurrency limits.
3. Trim unnecessary context to avoid repeatedly sending large blocks of irrelevant content.
4. For routine low-latency tasks, consider deepseek-v4-flash first.
5. For long-text or complex tasks, consider kimi-k3 -- run quality and cost tests before production use.
6. Review usage records regularly and analyze fees by model, group, user, and Key.

6. Exporting Usage Records

Fee details may include account, model, call time, and IP information. Before exporting a report, confirm the intended recipients and usage scope. When sharing externally, remove API Keys, personal information, and business content.

7. FAQ

Q: Why does my actual fee differ from this article? A: This article uses demonstration unit prices. Your actual fee depends on the model, group, standard price, discount, actual token count, and whether the cache was hit at the time of the call.

Q: Which is cheaper, deepseek-v4-flash or kimi-k3? A: Refer to the current prices in the model plaza and compare holistically, factoring in output length, task quality, and retry count.

Q: Do all models support caching? A: No. Whether a model supports caching and the cache unit price are indicated in the model plaza and fee details.

Q: If prices are adjusted later, will historical records change? A: No. Historical records are saved with the price and discount in effect at the time of the call, ensuring reliable reconciliation.