Due to Byte-Pair Encoding (BPE) character splits, German text generates 1.25× more tokens than equivalent English text. 1,000 German words consume ~1,666 tokens in Gemini 3.5 Flash-Lite.
Relative to English baseline (1.0×).
Price to send 1,000 words of German text into context.
Discounted price for repeated German system context.
| Word Count Scale | German Tokens | German Input Cost | English Equivalent Cost | Tokenization Penalty |
|---|---|---|---|---|
| 1,000 words (Short Article) | 1,666 | $0.0005 | $0.0004 | +$0.00010 |
| 10,000 words (Whitepaper / Report) | 16,663 | $0.004999 | $0.003999 | +$0.001 |
| 50,000 words (Book / Corpus) | 83,313 | $0.025 | $0.02 | +$0.004999 |
| 100,000 words (Enterprise Repository) | 166,625 | $0.05 | $0.04 | +$0.009998 |
Very long compound words are broken into 3-4 subword tokens.
Modern BPE vocabularies handle standard German compound stems efficiently.
Most LLM tokenizers are primarily trained on English-heavy web datasets. Non-Latin characters in German (Latin) split across multiple Byte-Pair Encoding (BPE) sub-word or multi-byte UTF-8 tokens, requiring approximately 1.25× more tokens to encode the exact same semantic meaning as English.
Modern BPE vocabularies handle standard German compound stems efficiently. In addition, enabling prompt caching on static German instructions or documentation saves 75–90% on input token rates.
Yes. Gemini 3.5 Flash-Lite has strong multilingual comprehension and generation capabilities in German (Deutsch). The difference is purely computational and financial due to sub-word token splits.