★ DeepSeek · token counter
DeepSeek Token Counter
DeepSeek's API docs publish a simple conversion rule for English and Chinese text. This page applies that rule, labels the result as an estimate, and shows the input cost per call on DeepSeek's current models.
- Runs in your browser
- No signup
- 306
- Characters
- 54
- Words
- —
- Tokens / word
Tokens per word uses the o200k_base count. Characters count emoji and accented letters as one each.
Tokens
92
Estimated with DeepSeek's published rule: ~0.3 per English and ~0.6 per Chinese character.
Context window used
—
Context window not published for this model.
Input cost per call
$0.00003
Per 1,000 calls
$0.0276
Every model, same text
| Model | Tokens | Accuracy | Input / call |
|---|---|---|---|
| Claude Fable 5.1Anthropic | … | ≈ Estimate | — |
| Claude Opus 5.5Anthropic | … | ≈ Estimate | — |
| Claude Sonnet 5.5Anthropic | … | ≈ Estimate | — |
| Claude Haiku 5.5Anthropic | … | ≈ Estimate | — |
| GPT-6 AstraOpenAI | … | ✓ Exact | — |
| GPT-6.1 SolOpenAI | … | ✓ Exact | — |
| GPT-6 LunaOpenAI | … | ✓ Exact | — |
| Gemini 4 ArgonGoogle | 77 | ≈ Estimate | $0.00015 |
| Gemini 3.8 FlashGoogle | 77 | ≈ Estimate | $0.00006 |
| Gemini 3.1 Flash-LiteGoogle | 77 | ≈ Estimate | $0.00002 |
| Mistral Large 4Mistral | … | ≈ Estimate | — |
| Mistral Small 4Mistral | … | ≈ Estimate | — |
| DeepSeek V4.1 FlashDeepSeek | 92 | ≈ Estimate | $0.00003 |
| DeepSeek V4 ProDeepSeek | 92 | ≈ Estimate | $0.00012 |
| Llama (open weights)Meta | … | ≈ Estimate | — |
OpenAI rows use o200k_base, exact for that tokenizer. tiktoken hasn't mapped GPT-6 model names yet, so OpenAI hasn't confirmed GPT-6 uses it. Input cost is list price for your text alone, before system prompts, caching or output.
See the tokens
o200k_base
Loading the tokenizer…
92 tokens on DeepSeek V4.1 Flash.
Prompts eating your context window? I build RAG and prompt pipelines that use a fraction of the tokens.
How DeepSeek tokenizes
DeepSeek's docs give approximate ratios: about 0.3 tokens per English character and about 0.6 tokens per Chinese character, and they note the ratio varies by model.
The real count for a request is in the usage field of the API response, which is what you are billed on.
Exact or estimate?
Estimate. We apply DeepSeek's published rule: 0.6 tokens per Chinese (CJK) character and 0.3 tokens per other character.
The rule is designed for English and Chinese. Code and other languages may land further from the real count.
Current DeepSeek models
Context windows and input prices from the pricing data used across these tools.
| Model | Context window | Input / 1M |
|---|---|---|
| DeepSeek V4.1 Flash | Not published | $0.30 |
| DeepSeek V4 Pro | Not published | $1.32 |
Quirks worth knowing
- Chinese text costs roughly twice as many tokens per character as English under DeepSeek's rule.
- Peak and off-peak pricing differs on DeepSeek. The cost column uses the peak rate from the pricing data.
- Context windows for the current DeepSeek models were not confirmed from a primary source, so the meter shows "not published".
Questions people actually ask
How does DeepSeek count tokens?
Is this DeepSeek count exact?
Why does Chinese text use more tokens?
Which counts on this page are exact?
Is my text sent anywhere?
Count by model family
- Claude token counter
- GPT token counter
- Gemini token counter
- Llama token counter
- Mistral token counter
Or compare every family at once on the main token counter.
More from the workbench
Fine print
Tools: on the house.
Engineer: by the hour.
Counting tokens was free. Spending fewer of them is a project.
I'm Himanshu, an AI and full-stack engineer. I build MVPs, AI agents and the growth plumbing around them for early-stage startups.