On this page

Use returned usage as the source of truth

DeepSeek explains that token estimates vary by language and tokenizer. The actual usage returned by the API is the number to trust for a completed request. Capture input and output tokens, then apply the current model price when you need a cost estimate.

If your client uses caching, separate cached and uncached input where the response makes that distinction. A high cache hit rate can materially change cost without changing the visible prompt length.

Watch account-level spend

Per-request logs explain individual calls. The platform account explains the bigger picture: funded balance, amount spent, recent tokens, request count, and activity by API key or model.

Use both views. A local log can miss requests from another machine, while an account total cannot tell you which application created a specific spike.

Keep your usage visible while you work. Get Super Notchy for $27 ↗

Set a review rhythm

Check the balance before a batch job, review usage after the batch, and set an alert before the balance becomes operationally risky. Super Notchy can show DeepSeek platform totals after an explicit sign-in inside the app, so the account status stays visible beside subscription-based assistants.

Avoid publishing a fixed cost estimate in your workflow. Model prices and plan terms change. Read current pricing, then calculate from the actual usage fields.

Sources and further reading