Gemini 4 Argon ships with a new tokenizer. The same text no longer gives the same token count as on Gemini 3.x, and token counts are what you pay for and what fills the context window. So I measured what changed, in 41 languages.
The short version:
- Korean needs 8% fewer tokens and Japanese 7% fewer for the same text. English did not change by a single token.
- Indian languages went the other way: Hindi, Tamil and Kannada need 12% to 19% more.
- Added up over all 41 languages the change is only +0.7%. The tokenizer did not get better or worse overall; the cost moved between languages.
The test
- 10 everyday requests (write a function, plan a trip, a product description, meeting notes and so on), translated into 41 languages: 410 texts
- Counted in the Google AI Studio token counter on Gemini 4 Argon and Gemini 3.8 Flash, 2026-10-21. Gemini 3.8 Flash and 3.1 Pro gave identical counts in an earlier test, so I call them 3.x
- The counter adds a fixed amount, which I subtracted: on Argon the one-letter text "a" shows 32, so 31; on 3.8 Flash it shows 2, so 1
- GPT (o200k) and Claude Haiku 5.5 numbers are from the same 410 texts, measured earlier
Result 1: the same 10 texts, before and after
| Language | Gemini 3.x | Gemini 4 Argon | Change |
|---|---|---|---|
| English | 373 | 373 | 0% |
| Chinese (Simplified) | 360 | 354 | -2% |
| Japanese | 417 | 387 | -7% |
| Korean | 498 | 456 | -8% |
| Hindi | 470 | 528 | +12% |
| Kannada | 679 | 794 | +17% |
| Tamil | 568 | 675 | +19% |
What matters is that the direction is consistent. English was identical on all 10 texts. Korean and Japanese went down on all 10 (-3% to -15% per text), and Hindi, Kannada and Tamil went up on all 10 (+6% to +31% per text). This is the tokenizer, not one odd sentence.
In characters it reads like this: one Korean token now holds 1.94 characters instead of 1.78, so the same context window fits about 9% more Korean text. One Kannada token went from 2.71 characters to 2.32.
Result 2: group by script and a pattern appears
All 41 languages (one request text each), grouped by writing system. Numbers are the median change from Gemini 3.x.
| Script | Languages | Median change | Range |
|---|---|---|---|
| Indic scripts (Hindi, Bengali, Tamil, ...) | 10 | +17% | -8% to +24% |
| Chinese, Japanese, Korean | 4 | -6% | -15% to -3% |
| Thai | 1 | -9% | |
| Arabic, Persian, Hebrew | 3 | -5% | -6% to -3% |
| Latin (English, Spanish, ...) | 20 | -2% | -10% to +3% |
| Cyrillic, Greek | 3 | 0% |
9 of the 10 Indic-script languages went up (the exception is Punjabi, -8%), and all four CJK languages went down. Among Latin-script languages, English, German and French stayed put, while languages with long, heavily inflected words dropped: Finnish -10%, Hungarian -8%, Polish -7%. Russian, Ukrainian and Greek did not change by a single token.
What it suggests: same total, different split
Summed over the 41 texts, Gemini 3.x needs 2,069 tokens and Argon 2,083, a difference of +0.7%. The median "language tax" (extra tokens versus English) is +41% on both.
A tokenizer splits text using a vocabulary of fixed size. Languages whose common character sequences get more entries in that vocabulary need fewer tokens. These results look like what you would expect if part of the vocabulary moved from Indic scripts to CJK and some European languages. Google has not described the change, so this is an inference from the measurements, not a confirmed fact.
Gemini's edge in Indian languages mostly went away
Gemini 3.x was unusually efficient for Indian languages. Much of that advantage is gone in Argon. Tokens for the one request text:
| Language | GPT (o200k) | Gemini 3.x | Gemini 4 Argon |
|---|---|---|---|
| Hindi | 51 | 41 (-20% vs GPT) | 48 (-6%) |
| Bengali | 57 | 39 (-32%) | 46 (-19%) |
| Tamil | 67 | 52 (-22%) | 63 (-6%) |
| Kannada | 61 | 58 (-5%) | 72 (+18%) |
| Malayalam | 63 | 65 (+3%) | 71 (+13%) |
Hindi went from 20% fewer tokens than GPT to 6% fewer, and Kannada and Malayalam now need more than GPT.
Next to the other models
Total tokens for the same 10 texts:
| Language | Gemini 4 Argon | GPT (o200k) | Claude Haiku 5.5 |
|---|---|---|---|
| English | 373 | 350 | 520 |
| Korean | 456 | 534 | 935 |
| Japanese | 387 | 570 | 684 |
| Hindi | 528 | 565 | 1,223 |
| Kannada | 794 | 679 | 3,276 |
In English GPT needs 6% fewer tokens than Argon, but in Korean the order flips: Argon needs 15% fewer than GPT and 51% fewer than Claude. Fewer tokens is not automatically cheaper, since the price per token differs by model.
What changes in practice
- Korean or Japanese on Gemini: at the same price per token, input costs drop 7% to 8%, and the same context window holds more text. Token-based limits and free tiers last correspondingly longer.
- Indian-language products moving from 3.x to Argon: budget for 10% to 20% more input tokens on the same traffic. For Kannada and Malayalam, GPT now needs fewer tokens.
- English-only products: no change in token counts.
- Cost estimators: rules of thumb like "characters divided by a constant" tuned on 3.x no longer hold for Korean or Indian languages. Measure per model.
Limits
- Input tokens only. Output very likely uses the same tokenizer, but I did not check it
- The 10-text comparison with Gemini 3.x covers 7 languages. The 41-language comparison uses one text, so single languages can move by a few percentage points; on that text Korean showed a larger -15%
- AI Studio counter numbers; the API's countTokens may differ by a few tokens
- The vocabulary explanation is an inference from the data, not something Google has confirmed
All 41 languages, chart and CSV (CC BY 4.0): https://tokensave.app/data/language-tax-index
Count your own text: https://tokensave.app
