Hindi costs 1.1× to 4.9× the tokens of English, by tokenizer

THE SHORT ANSWER

Hindi in Devanagari cost 4.55–4.91× English on Qwen, Mistral and OpenAI's cl100k_base, 1.61× on o200k_base, and only 1.13–1.30× on Indic tokenizers.
Dev-roman Hinglish (Devanagari with English technical terms) used about a third fewer tokens than pure Devanagari on Qwen: 623 against 929.
Small sample: 12 sentences. The script and the raw numbers are published below.

Why did we measure this?

Our AI agents write CBSE study material, much of it in Hindi, and our working rule said Devanagari costs “about 3×” the tokens of English. That number was an estimate. Tokens decide what an agent can read in one session, what an API call costs, and how much a small on-device model can hold, so we wanted the real figure, per tokenizer.

A tokenizer is the part of a language model that cuts text into pieces (tokens) before the model sees it. Every model family has its own, and they treat scripts very differently.

How did we test it?

We wrote 12 original school science and social-science sentences (Class 9–10 level) and wrote each one four ways. Same meaning, four scripts and styles:

VERSION	EXAMPLE (FIRST SENTENCE)
English	Plants make their own food from sunlight, water and carbon dioxide through photosynthesis.
Hindi, Devanagari	पौधे प्रकाश संश्लेषण के ज़रिए सूरज की रोशनी, पानी और कार्बन डाइऑक्साइड से अपना खाना खुद बनाते हैं।
Hinglish, Roman script	Plants photosynthesis ke zariye sunlight, paani aur carbon dioxide se apna food khud banate hain.
Dev-roman (Devanagari + English terms)	Plants photosynthesis के ज़रिए sunlight, पानी और carbon dioxide से अपना food खुद बनाते हैं।

We counted tokens with no special tokens added, using each tokenizer's own files (no model weights). Every Hugging Face count was checked against a second implementation, and we re-ran the whole script and got identical numbers. Two public parallel sets (FLORES+ and IN22-Gen) and two tokenizers (Gemma 3, Llama 3.2) are gated and were skipped.

How many more tokens does each version cost?

Each cell is the version's token count divided by the English count for the same tokenizer (1.00× = as cheap as English).

TOKENIZER	ENGLISH TOKENS	HINDI (DEVANAGARI)	HINGLISH (ROMAN)	DEV-ROMAN
OpenAI o200k_base	199	1.61×	1.48×	1.25×
OpenAI cl100k_base	202	4.91×	1.62×	3.25×
Qwen2.5-0.5B / Qwen3-0.6B	204	4.55×	1.61×	3.05×
Mistral-7B-v0.1	215	4.85×	1.68×	3.34×
Sarvam-1	211	1.30×	1.92×	1.18×
IndicBART (IndicBARTSS)	257	1.13×	1.74×	1.15×
mBERT (reference)	211	1.95×	1.50×	1.39×

Does dev-roman Hinglish save tokens?

On general-purpose tokenizers, yes, and by a lot. Keeping the technical terms in English, the way people in India actually speak, cuts the tokens for the same Hindi sentence by about a third on Qwen and older OpenAI tokenizers:

TOKENIZER	DEVANAGARI	DEV-ROMAN	CHANGE
OpenAI cl100k_base	991	657	−34%
Qwen2.5 / Qwen3	929	623	−33%
Mistral-7B-v0.1	1,042	719	−31%
mBERT (reference)	411	293	−29%
OpenAI o200k_base	321	249	−22%
Sarvam-1	274	248	−9%
IndicBART	291	296	+2%

Indic tokenizers already handle Devanagari well, so there the saving almost disappears. And Roman-script Hinglish, the cheapest option on general tokenizers, is the most expensive on Indic ones (1.92× on Sarvam-1).

What does this mean for building in Hindi?

Our “about 3×” rule was right only for dev-roman on older tokenizers (3.05–3.34×). Pure Devanagari on those tokenizers costs 4.5–4.9×; on o200k_base it costs 1.6×.
For our Hinglish small model, which starts from a Qwen base, dev-roman output needs about a third fewer tokens than pure Devanagari. That means more text per phone-sized context window, and faster output.
Pick the tokenizer with the language. For mostly-Devanagari text, an Indic tokenizer is close to English cost. For mixed text, a modern general tokenizer such as o200k_base is cheap across all four styles.

What does this not prove?

It is a small sample: 12 sentences, about 200 English tokens per tokenizer. Treat the ratios as indicative, not exact.
One person wrote all four versions, and Hinglish spelling has no standard; different wording would move the numbers.
Token cost says nothing about how well a model understands or writes each version.
A ratio is not a price; API prices per token differ by provider and model.
The public sets (FLORES+, IN22-Gen) and the Gemma and Llama tokenizers were not tested because they are gated.

What we decided

Our Hinglish small model targets dev-roman output. Before choosing any model for Hindi text, we measure its tokenizer on our own sentences first.

Questions

How many more tokens does Hindi cost than English?

On our 12-sentence school set, Hindi in Devanagari cost 1.61 times the English tokens with OpenAI's o200k_base tokenizer, and 4.55 to 4.91 times with cl100k_base, Qwen2.5, Qwen3 and Mistral. Indic tokenizers were close to English: 1.30 times for Sarvam-1 and 1.13 times for IndicBART.

Is it cheaper to write Hinglish in Roman script?

On general-purpose tokenizers, yes: Roman-script Hinglish cost 1.48 to 1.68 times English. On Indic tokenizers it was the most expensive version, 1.92 times on Sarvam-1 and 1.74 times on IndicBART, because they are built for Devanagari.

What is dev-roman Hinglish and why does it save tokens?

Dev-roman is Hindi written in Devanagari with technical terms kept in English, for example photosynthesis or sunlight. English words are cheap for most tokenizers, so on Qwen2.5 and Qwen3 dev-roman cost 623 tokens against 929 for pure Devanagari, about a third fewer.

RAW DATA

Script and data: bench.py · the 12 sentences · results (JSON) · results (table). The 12 sentences are ours, released under CC BY 4.0: reuse them with credit to Bodham Labs.

CITE THIS NOTE

Chandra, P. (2026). Hindi costs 1.1× to 4.9× the tokens of English, by tokenizer. Bodham Labs Notes.

BACKGROUND READING

Prabhakar Chandra

Founder of Bodham Labs. BTech in Electrical Engineering, MCA and B.Ed; teaches AI and builds the systems in our work ledger.