DeepSeek Token Counter

Count tokens for DeepSeek-V3 and DeepSeek-R1 with cost estimates. DeepSeek uses a custom tokenizer (closest comparable: GPT-3.5's cl100k_base). At $0.27/M input ($0.07 cached), DeepSeek-V3 is one of the cheapest near-GPT-4-class models on the market in 2026.

How to use the DeepSeek Token Counter

Paste text, pick V3 or R1, enter expected output and call volume. Cache hit ratio applies the 74-75% discount on cached input tokens. R1 reasoning tokens count toward output — a single R1 call can produce thousands of reasoning tokens before the visible answer.

DeepSeek pricing reality

DeepSeek-V3 at $0.27/M input / $1.10/M output is roughly 10x cheaper than GPT-4o on input and 10x cheaper on output. Quality is competitive with GPT-4 (not GPT-5) on most tasks. R1 is the reasoning variant — costs $0.55/$2.19 per million, comparable to GPT-5 Mini but with stronger STEM benchmarks.

Caching: input tokens matching a cached prefix cost $0.07/M (V3) or $0.14/M (R1) — a 74-75% discount. Useful when your system prompt is stable. The cache TTL is shorter than Anthropic's (typically 30-60 min for the official API).