DeepSeek token counter
DeepSeek has released its tokenizers with its open model weights, so exact counting can run entirely in your browser. Nothing is uploaded and nothing is estimated.
Why count DeepSeek tokens at all
DeepSeek's prices are low enough that many teams stop thinking about token counts, which is exactly how caching discounts go unclaimed. On the current V4.1 Flash model, cache-hit input bills at a fiftieth of the standard input rate, so knowing how many of your tokens sit in a repeated prefix is worth real money at volume. The pricing table carries the current numbers, and the prompt caching guide explains how prefix caching works mechanically.
DeepSeek off-peak hours
Every DeepSeek API rate is half price outside peak hours. Peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, which lands on US evenings and nights from Sunday to Thursday. The full schedule for Pacific, Mountain, Central and Eastern time, with a week chart in your own time zone, is on the DeepSeek off-peak hours page.
Counting notes
- The tokenizer is the reference implementation from DeepSeek's own releases, not an approximation. Coverage for the V4 generation is verified as part of the counter rollout.
- The current V4.1 Flash and V4 Pro models take 1,000,000 tokens of context with up to 384,000 output. See how that compares at context windows.
- Methods and verification: how we count.