DeepSeek V4.1 Flash: Calculate Cost per Accepted Task

DeepSeek pricing depends on the tokens billed, caching, and the applicable rate period. A low price for one call can still produce an expensive finished task when several attempts fail or require substantial editing.

Rate snapshot: September 28, 2026

The official DeepSeek pricing page listed the following USD charges for deepseek-flash, mapped to V4.1 Flash, during this review. Rates are per million tokens. Reopen the provider's table before estimating a new job or scheduling production traffic.

DeepSeek Flash API rates recorded on September 28, 2026

Token category	Off-peak USD / million	Peak USD / million
Input, cache hit	$0.003	$0.006
Input, cache miss	$0.15	$0.30
Output	$0.60	$1.20

The recorded peak windows were 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Use the provider's billing rules for the actual request. Your local workday or timezone may not match those windows.

Calculate a request with cached and uncached input

Suppose a fictional off-peak request bills 80,000 cache-hit input tokens, 20,000 cache-miss input tokens, and 10,000 output tokens. The arithmetic is:

Cached input:    80,000 / 1,000,000 * $0.003 = $0.00024
Uncached input:  20,000 / 1,000,000 * $0.15  = $0.00300
Output:         10,000 / 1,000,000 * $0.60  = $0.00600
Request total:                               $0.00924

At the recorded peak rates, the same billed token counts total $0.01848. These are illustrative calculations, not a measured job or a quote. Use the usage categories returned by your provider; do not assume repeated text qualifies as a cache hit.

Count every attempt in the finished task

A rejected output still belongs in the task's cost. Keep an attempt log with model identifier, rate period, billed input categories, output tokens, external tool charges, and acceptance result. Group related retries under one task identifier so failures do not disappear from the budget.

For example, if ten attempts cost $0.10 in total and only eight outputs meet your acceptance rules, the model cost per accepted output is $0.0125. With zero accepted outputs the calculation has no usable denominator: record a failed evaluation. Do not report the cost as zero.

Keep review time in a separate column

Record minutes spent reading, correcting, and checking the answer. You can convert them to a cost using an explicit hourly assumption, but keep the raw minutes too. A reviewer may reject a fluent answer quickly while taking much longer to repair a subtle factual error.

Choose acceptance rules before comparing models. For a data extraction job, they might require every invoice number, amount, and currency to match the input and every missing field to remain blank. For code, they might require a reproduced failure, a bounded fix, and passing checks for behavior that must remain unchanged.

Choose a realistic workload

Use a sample that includes short and long inputs, repeated context, and examples that previously failed. Keep the prompt, source files, time limit, and tools consistent. Report first-attempt success separately from success after retries. Change one variable when investigating a cost increase.

Save failed examples and rerun them after a model alias or application changes. An alias can be convenient, but a familiar name does not establish identical behavior over time. If the provider offers version controls appropriate to your workflow, record them alongside the model identifier.

What the token bill leaves out

Tool calls, a gateway's charges, storage, and human review may sit outside the model's token price. Local hosting has hardware and operating costs of its own. Keep currencies consistent and specify which costs your comparison includes. A low token rate is useful only when the accepted output meets the required quality.

A worksheet to keep

Task ID:
Model and provider:
Date and applicable rate period:
Cache-hit input tokens:
Cache-miss input tokens:
Output tokens:
Attempts and external charges:
Acceptance checks passed / failed:
Review and correction minutes:
Total spend / accepted outputs:

Use the resulting record to decide which workload belongs on which model. If review time dominates, improve the task specification and evidence checks before chasing a smaller token rate.