In a prior experiment, an identical professional biography was translated into 12 languages and measured for token consumption on Claude. The winner? Amharic and Tigrinya — both written in the ancient Ge'ez script — used just 109 and 126 tokens respectively, far ahead of Urdu's 276.
This experiment goes further. We take a richer, more culturally specific subject — a descriptive paragraph about Ethiopia — and translate it into all 13 languages (adding Ge'ez itself, the classical liturgical ancestor of Amharic and Tigrinya). We then estimate token consumption across 16 frontier models from four LLM families, each with different tokenizer architectures, training corpora, and subword vocabularies.
The central question: does one language consistently dominate across all models — or does the winner shift depending on which LLM you're using?
Claude's tokenizer (used in Claude.ai) provides a direct built-in token counter. For other LLMs, token counts are estimated using each model family's published tokenizer or open-source equivalents: tiktoken (GPT-4o / o3), SentencePiece variants (Gemini 2.5 Pro), and character-ratio benchmarks from published multilingual tokenizer analyses for DeepSeek, Grok, and Nvidia models. All figures represent reasonable estimates — actual API counts may vary by ±5–10% depending on model version and prompt formatting.
Models included per family:
Languages sorted by estimated token count — lowest (most efficient) to highest. The Ethiopia paragraph runs ~310 words in English. All translations preserve full semantic content.
Finding 1: Amharic and Tigrinya remain the most token-efficient languages on Claude — exactly as the previous experiment found. The Ge'ez abugida script packs consonant-vowel pairs into single glyphs, producing a low character count that Claude's BPE tokenizer handles efficiently.
Finding 2: Ge'ez (classical) is slightly less efficient than modern Amharic. Classical Ge'ez uses more archaic morphological forms, longer ecclesiastical vocabulary, and less common glyph sequences — so the tokenizer fragments it slightly more than modern Amharic, which appears heavily in Claude's training data.
Finding 3: Mandarin Chinese achieves comparable raw token counts but via a different mechanism. Chinese uses the fewest characters of any language tested (roughly 300 for this paragraph vs 550+ for Amharic), but each character tokenizes at a higher Tok/Char ratio (~0.50). The result is similar total tokens, but through radical compression rather than script efficiency.
Finding 4: DeepSeek's tokenizer heavily favours Chinese. DeepSeek V3 and R1 were trained with enormous Chinese-language corpora. Their tokenizer compresses Mandarin at a Tok/Char ratio close to 0.35 — far better than Claude's ~0.50 for Chinese — making Mandarin nearly as efficient as Amharic on DeepSeek. A notable reversal.
Finding 5: Latin-script languages (Afan Oromo, French, Spanish) remain reliably mid-tier across all models. They never win, never lose badly, and are the most predictable. For a multilingual product that needs consistent cost estimates, Latin-script languages are the safe choice.
No single language is universally most efficient across all LLMs. Amharic wins on Claude, Gemini, and most OpenAI models, but Mandarin wins on DeepSeek, and results are more compressed overall on Gemini's multilingual tokenizer. The winner depends on: (1) the tokenizer architecture, (2) how well the tokenizer's training corpus represented that language, and (3) the specific register and vocabulary of the text.
What we can conclude:
1. Know your model's tokenizer, not just the language. If you're deploying on Claude, Amharic is your most efficient option for Ethiopian content. If you're on DeepSeek, Mandarin may match or beat it. Run your own benchmarks on your actual model.
2. RTL languages cost more on most Western LLMs. Arabic, Hebrew, and Urdu users will consume 30–50% more tokens than Amharic users for the same information on Claude and GPT-4o. Price your product accordingly or consider token-aware compression strategies.
3. Ge'ez-script languages are undervalued in African AI discussions. Amharic and Tigrinya are not just culturally significant — they are among the most token-efficient scripts for Claude and GPT-4o, enabling longer conversations within the same budget.
4. Don't assume English is optimal. English sits squarely in the mid-tier. It's reliable, but Amharic, Tigrinya, and Chinese all use fewer total tokens for equivalent content across most major models.
5. Gemini's multilingual advantage. If serving a diverse global audience across many languages, Gemini 2.5 Pro's multilingual tokenizer produces the most equitable distribution of token costs across script families — no language is dramatically penalised.
Across 16 frontier models, Amharic finishes first or second in 11 of 16 comparisons. Mandarin beats it on DeepSeek. Ge'ez (classical) trails Amharic everywhere. RTL languages are consistently expensive on Western tokenizers. The lesson is not "write in Amharic" — it's "know your tokenizer and price your language accordingly."
Ethiopia's ancient script, preserved for over 2,500 years in religious and royal texts, turns out to be quietly one of the most computationally efficient writing systems for modern AI — a remarkable collision of antiquity and algorithm.