I’m trying to estimate the cost of using DeepSeek for a small production workload, and the simple “requests per day” calculation feels misleading. Some inputs are short, others include a lot of context, and a few tasks ask for much longer responses.

When you estimate DeepSeek API costs, what usage details make the biggest difference? Do you track input and output tokens separately, account for retries, or measure a representative sample before projecting monthly spend? I’m interested in a practical method that doesn’t assume every request looks alike.