Research · measured 18 October 2026

What Indian languages cost on AI models.

We sent the same short passage, in English, Hindi, Marathi, Bengali, Tamil and Telugu, to 11 AI models through Rapid Router and recorded how many tokens each provider billed for it. Tokens are what you pay for, so this is how much more the same message costs in an Indian language.

The same passage in six languages on 11 models: how many more tokens, and rupees, an Indian language costs.

2.5×

Hindi tokens versus English, averaged over 11 models

1.3×

Most efficient for Indian languages: Sarvam-105B

6.2×

Least efficient: GLM-5.3

Tokens per language, as a multiple of English

Darker means more tokens for the same meaning. Sorted from most to least efficient across the five Indian languages.

Model	English tokens	Hindi	Marathi	Bengali	Tamil	Telugu	Average
Sarvam-105BSarvam	123	1.2×	1.4×	1.1×	1.4×	1.4×	1.3×
Gemini 3.6 FlashGoogle	123	1.2×	1.4×	1.1×	1.4×	1.8×	1.4×
GPT-5 MiniOpenAI	121	1.4×	1.8×	1.5×	2.1×	2.0×	1.8×
Ministral 3 14BMistral	124	1.7×	2.1×	1.8×	2.4×	2.3×	2.1×
Grok 4.6xAI	121	1.6×	2.0×	2.1×	2.8×	4.3×	2.6×
DeepSeek V3.2DeepSeek	124	2.8×	2.9×	1.8×	3.3×	3.9×	2.9×
Kimi-K2-Instruct-0905Moonshot	121	2.8×	2.9×	2.8×	4.5×	4.2×	3.5×
Claude Haiku 4.5 (latest)Anthropic	127	3.2×	3.6×	3.7×	4.8×	4.4×	3.9×
Llama-3.3-70B-InstructMeta	123	2.4×	2.8×	5.6×	8.3×	8.9×	5.6×
Qwen3 235B A22B Instruct 2507Qwen	123	4.6×	4.7×	4.9×	6.8×	7.7×	5.7×
GLM-5.3Z.ai	123	4.3×	4.7×	4.7×	8.3×	8.9×	6.2×
Average across models		2.5×	2.8×	2.8×	4.2×	4.5×

What it does to the bill

The input cost of 10,000 customer messages like the test passage, in English and in Hindi, at each model's live price on Rapid Router. Output tokens grow the same way, so a reply in Hindi costs more again.

Model	Price per 1M input	English	Hindi	Tamil
Sarvam-105B	₹35.37	₹43.51	₹54.12	₹62.25
Gemini 3.6 Flash	₹73.69	₹90.64	₹112.74	₹129.69
GPT-5 Mini	₹24.56	₹29.72	₹42.49	₹61.65
Ministral 3 14B	₹19.65	₹24.37	₹40.87	₹57.77
Grok 4.6	₹197	₹237.77	₹389.07	₹654.35
DeepSeek V3.2	₹25.55	₹31.68	₹87.62	₹103.46
Kimi-K2-Instruct-0905	₹73.69	₹89.16	₹250.54	₹403.81
Claude Haiku 4.5 (latest)	₹98.25	₹124.78	₹404.79	₹596.38
Llama-3.3-70B-Instruct	₹9.83	₹12.08	₹28.59	₹100.31
Qwen3 235B A22B Instruct 2507	₹8.84	₹10.88	₹49.61	₹74.45
GLM-5.3	₹138	₹169.19	₹734.52	₹1,404.39

Method

One passage of 107 English words about a small shop using an AI assistant, translated by us into the five languages with the same meaning and register. Each version was sent once to each model through the Rapid Router API on 18 October 2026. We recorded the prompt token count each provider reported, minus the count for a one-word message, so the fixed overhead of the chat format cancels out.

Limits worth knowing: it is one passage, so treat the figures as indicative rather than exact; native writers phrase things differently from a translation; and tokenizers change between model versions, so a newer release can move a lot. Two models that share a tokenizer family, or that fall back to splitting rare scripts byte by byte, can report identical counts; that is a real result, not a copy.

A small clothing shop in Jaipur gets about two hundred customer messages every day. Most of them ask the same things: is this size available, how long will delivery take, can I return it if it does not fit. The owner used to answer every message herself late at night. Now an AI assistant reads each message, checks the stock list and the courier's tracking page, and writes a short, polite reply in the customer's own language. She reads the replies once in the morning and corrects the few that are wrong. Her customers get answers in minutes instead of hours, and she has her evenings back.

For models built for Indian languages, see the Indic collection; for quality rather than cost, the best model for each task.

Common questions

Why do Indian languages cost more on AI models?

Models bill per token, and a tokenizer trained mostly on English splits Devanagari, Bengali, Tamil and Telugu text into many more pieces than the same meaning in English. More tokens for the same message means a bigger bill and less room in the context window.

Which AI model is cheapest for Hindi?

By token count alone, Sarvam-105B came out most efficient across the five Indian languages we measured (1.3× English on average). Whether it is cheapest also depends on its price per token; the cost table on this page multiplies the two.

Does this mean Hindi answers are worse?

No. This study measures tokens, not quality. A model can tokenize Hindi expensively and still answer well, or the reverse. Quality rankings are on our best-model pages.

Can I use these numbers?

Yes, under CC BY 4.0. Cite "Indian-language token study, Rapid Router, 18 October 2026 (rapidrouter.ai/research/indic-token-cost)".