What GoTo’s regional-language AI reveals about language costs and reasoning strengths

The same request can occupy different amounts of an AI system’s context in Japanese and English. Asking a question in Japanese and asking the model to generate its reasoning in Japanese can also affect performance differently. Understanding language-dependent usability requires keeping those issues separate.

GoTo’s released Gemma 3 derivative provides a concrete example. TechNode reported on the model on 19 October 2026; that date identifies the report. Its model card lists English and five languages used in Indonesia and says it retains the default Gemma 3 tokenizer. Better regional-language scores therefore do not imply shorter token sequences for identical text. Text segmentation and the ability to use that text deserve separate evaluation.

Four different meanings of language in an AI system

Claims that an AI “thinks in English” can blur at least four things: the user’s input, reasoning text generated along the way, internal representations during computation, and the final answer. A workflow could accept a Japanese question, generate English reasoning text, and deliver a Japanese response.

Chain of Thought (CoT) refers here to reasoning text that a model generates while solving a problem. Generated text becomes part of the context for subsequent generation, so changing its language can affect the result. Internal representations, by contrast, are numerical vectors computed across layers. Observing intermediate prose and inspecting internal computation are different activities.

The distinction changes how results should be read. Strong performance on Japanese questions does not guarantee reliable Japanese reasoning text. Fluent Japanese explanations do not establish that every internal computation corresponds to Japanese. Evaluation should reflect whether the workflow needs correct answers, explanations readable in the user’s language, or both.

What the GoTo results show

The model is the 27B GoToCompany/gemma3-27b-cpt-gotoai-v3-instruct. Its six listed languages are English, Indonesian, Javanese, Sundanese, Batak Toba, and Balinese. The model card advertises a 128K-token context window, with weights available under the Gemma license.

The comparison is with Google’s instruction-tuned Gemma-3-27B-IT. The table uses the official leaderboard’s JSONL data to compare models within each task grouping. These are scores on a 0–100 scale, not simply percentages of questions answered correctly: the evaluation includes normalization for chance performance and aggregates multiple tasks. General capability evaluation uses native-language prompts in a zero-shot setting, with 100–1,000 instances per dataset.

The average covers the five languages other than English and excludes the separately listed IndoMMLU score. Indonesian and the regional languages have different task mixes. Javanese’s 71.09 and Indonesian’s 65.19 therefore cannot establish superior “thinking ability” in Javanese. The meaningful comparison is between the two models within a row.

This is developer-reported evaluation, not an independent replication, and higher aggregates do not mean every component improved. Indonesian question answering, for example, falls from 82.54 to 81.93. The repository’s mergekit configuration records a linear merge, but the inspected materials do not isolate continued pretraining, instruction tuning, and merging through separate ablations. The score difference belongs to the released model as a whole.

The same meaning can take different numbers of tokens

Tokens are the units into which a model’s tokenizer divides text. A character does not always correspond to one token: a token may represent a whole word, part of a word, or a smaller encoded piece of a character. The tokenizer determines those units and their identifiers, so the same text can have different lengths under different vocabularies and segmentation rules.

A NeurIPS 2023 study of tokenizer disparities compared approximately 2,000 FLORES-200 sentences and their human translations. Its comparison concerns the tokenized lengths of equivalent sentences, rather than texts with equal character counts or equal numbers of whitespace-separated words.

In the same study, mT5 produced ratios of 0.90 for Japanese and 0.92 for Chinese, both below English. A ratio such as 2.30 is therefore not an inherent constant of Japanese. These are historical measurements of particular tokenizers and corpora, not multipliers for a current ChatGPT bill.

The denominator matters too. The 2026 TokLens paper reports Japanese tokens per word at 56 times English for GPT-2, but “word” means a whitespace-delimited unit. Japanese normally lacks spaces between words, so one such unit can contain much more text. The authors explicitly warn that these units do not correspond to linguistic words in Japanese.

That 56-fold figure is not a claim that equivalent content costs 56 times as much. Checking whether the denominator treats a long Japanese sequence and a single English word as equivalent units changes the interpretation. For a cost comparison, matched-meaning inputs and an equivalent output-quality requirement are more relevant.

Tokenizer design changes the cost of a language

Tokenization disadvantages can change. In the examples accompanying the GPT-4o announcement, a Japanese greeting decreased from 37 tokens to 26; the Chinese example went from 34 to 24, and the English example from 27 to 24. These compare old and new tokenization within each displayed example. They are neither language-wide averages nor a controlled cross-language billing comparison.

The Gemma 3 Technical Report describes roughly 260,000 vocabulary entries (262k in its tokenizer section) and increased multilingual monolingual and parallel training data. A large vocabulary and multilingual data do not, by themselves, establish equal proficiency in every language.

GoTo retains the default Gemma 3 tokenizer. With identical text and encoding settings, regional adaptation alone does not shorten the input token sequence, yet evaluation scores can still change. Conversely, Tokenization Is More Than Compression tested 64 models from 350M to 2.4B parameters and found that minimizing sequence length under a fixed vocabulary did not reliably improve downstream performance.

Representing text with fewer units and understanding it well enough to solve a task are distinct improvement targets. For an application developer, the practical question is how much resource is required to maintain the needed quality, rather than whether either metric looks good in isolation.

Bills, computation, context, and cost per successful task

When saying that Japanese is more expensive, it helps to separate four quantities: the service bill, computing resources, the amount of information that fits in context, and the total cost of a successful task. They are related, but they need not change by the same factor.

For a token-billed API, start by applying the relevant rates separately to input and output. Distinguish cached input when it has its own rate, and include hidden reasoning tokens or tools where the service bills for them. A flat-rate chat subscription and a usage-billed API also require different accounting. A real comparison needs the chosen provider’s billing definitions for that model.

For an illustrative calculation, assume input costs ¥1 per 1,000 tokens and output costs ¥5 per 1,000. An input of 1,000 plus output of 1,000 costs ¥1 + ¥5 = ¥6. Increasing only the input to 2,000 yields ¥2 + ¥5 = ¥7. These are hypothetical rates, not a real API price quote. Output is fixed; reasoning, caching, and tools are also held constant, with no additional charges in this example.

NVIDIA’s explanation of inference optimization distinguishes prefill, which processes the prompt, from decode, which generates successive tokens. Longer input affects prompt processing and memory; longer output adds sequential generation. Wall-clock time also depends on parallelism, memory bandwidth, caching, and implementation. Doubling input tokens does not establish a universal twofold or fourfold increase in total price, latency, or energy.

A 128K-token context window is not a limit expressed in semantic information or pages. If a translation consumes more tokens, less room remains for other material and generation. But if shortening the material forces repeated follow-up requests to restore omitted conditions, reduced initial input alone does not establish a more efficient workflow.

One practical measure is total API and tool spending across all attempts, plus translation and human checking or repair, divided by the number of accepted results. This is an editorial proposal for workflow accounting. A cheap call that repeatedly mishandles proper names may require substantial manual correction. Keep unit-price comparisons separate from the cost of completed work.

An experiment where English reasoning text performed better

Separately from token counts, the language of generated reasoning can affect performance. Understanding a Japanese question does not necessarily imply an equal ability to sustain long reasoning in Japanese. Studying that gap requires varying input language and reasoning language separately.

Table 1 of Long Chain-of-Thought Reasoning Across Languages evaluates a research checkpoint: Qwen3-8B-Base supervised-fine-tuned on 20,000 English reasoning examples. MATH-500 questions were translated with Gemini-2.0-Flash, with one generation per question, temperature 0.6, top-p 0.95, and a 32,768-token generation limit. These are not scores for the standard released Qwen3-8B model.

Scoring requires both a correct final answer and at least 90% of evaluated chunks in the specified language. The English-CoT condition also requires the answer in English. Thus, 93.2% versus 77.6% is not a comparison of mathematical correctness independent of language: it combines input, generated language, and the scoring rule.

Within this setup, Japanese and English inputs produce similar scores when generation is in English, while requiring Japanese reasoning reduces the score. That illustrates why input comprehension and sustained reasoning text in a language should not be treated as identical capabilities. Results from an English-fine-tuned model do not establish an inherent, universal superiority of English.

The gap changes with the model and its training

The size of the gap varies by model. In a separate study, Language Matters, MATH-500 was translated with Google Translate API, and short language-specific phrases were prefilled to steer reasoning generation: “Okay” in English, for example, or mazu (“first”) in Japanese. The following results use Japanese input, temperature 0.6, and top-p 0.95.

QwQ-32B’s 90.2% versus 87.4% is a much smaller gap than the Llama-8B derivative’s 74.6% versus 32.6%. These are separately trained models, so the comparison does not isolate scale. Translation and scoring also differ from the preceding study; its numbers should not be combined with these in a single model ranking.

The M-Thinker study examines training to sustain native-language reasoning. The comparison below runs from DeepSeek-R1-Distill-Qwen-7B to M-Thinker-7B Iter-2. Evaluation uses MMATH’s 374 questions from AIME24/25, CNMO, and MATH-500 across ten languages, with four runs and macro-averaging across subsets. Japanese is among the training languages.

Here, accuracy measures answer correctness regardless of output language. Language consistency separately measures whether reasoning and answer retain the input language. This differs from the preceding study’s language-compliance-gated score. Both Japanese measures improved, showing that training can improve correctness and language consistency together.

Naive-RL, which rewards correctness, nevertheless achieves a slightly higher mean accuracy: 65.43% versus 65.04%. Its Japanese accuracy is also higher, 66.11% versus 58.23%, while Japanese language consistency is 0%. This does not show that native-language reasoning is always more accurate; it exposes a tradeoff between answer accuracy and keeping the response in the user’s language.

Internal representations are not English prose

What happens beneath the visible reasoning text? One interpretability method, the logit lens, projects intermediate-layer vectors into the output vocabulary to examine which tokens are favored. It is a way of probing internal states, not evidence that a human-readable English sentence is literally stored inside the model.

Do Multilingual LLMs Think In English? probes Llama-3.1-70B, Gemma-2-27B, Aya-23-35B, and Mixtral-8x22B. It reports English-leaning intermediate representations for semantically loaded words even during generation in languages such as French. This supports examining shared semantic components separately from language-specific ones, within the limits of the models and methods studied.

A separate Japanese-model study compares English-centric Llama-2-13B, Swallow-13B continued-pretrained on Japanese, and LLM-jp-v2.0-13B, pretrained on a 50:50 mix of English and Japanese. Under Japanese-processing conditions, it observes a mixture of English and Japanese tendencies in Swallow and more Japanese-leaning representations in LLM-jp; the distributions change with the output language. This limited study uses 8-bit-quantized 13B-class models and 166 examples for translation, repetition, and cloze tasks. It does not represent all current Japanese models, but indicates that language-related internal tendencies can differ with training.

Anthropic’s analysis of Claude 3.5 Haiku examines shared conceptual circuits and language-specific circuits across questions in different languages. English-leaning representations and language-independent conceptual sharing can coexist. These studies inspect different layers, objects, and mechanisms; one finding need not cancel the other.

For users, the practical implication is to avoid overinterpreting language instructions. “Think in English and answer in Japanese” may influence generated text, but does not establish that every internal vector computation has switched to English. Evaluate the observable outputs and task results rather than assuming complete control over hidden computation.

When translation helps, and when the original matters

If English tokenization is shorter and English reasoning performs better, should everything be translated? Translation adds processing, latency, and its own input and output. Translating the answer back adds another stage. Any savings from English need to be compared against those additions across the complete task.

Is Translation All You Need? finds that translating into English can help some tasks with English-centric models, while native-language prompting can be more effective for culturally and linguistically nuanced tasks. Beyond English separately examines translation of instructions, context, examples, and output across 35 languages on question answering, natural language inference, named-entity recognition, and summarization. Translation is not limited to an all-or-nothing choice.

Language mixing during reasoning is not automatically a failure. A study of English–Chinese language mixing finds benefits from allowing mixing on MATH500, but conditions favoring Chinese-only generation on Gaokao Cloze, a Chinese college-entrance mathematics fill-in-the-blank task. The experiment restricts available languages by masking vocabulary, rather than merely changing a language instruction in a prompt. These results concern English–Chinese models and specific tasks; they do not directly establish the same effect for Japanese.

The language in which a benchmark was created also matters. MultiNRC uses native-authored reasoning involving wordplay, riddles, culture, and traditions instead of relying exclusively on translated English questions. It covers French, Spanish, and Chinese, not Japanese. The broader evaluation question remains relevant to Japanese: translated English problems may not cover all the capabilities local users need.

For a Japanese wordplay explanation, sound and spelling may themselves be part of the problem. The Japanese words for bridge, chopsticks, and edge share the reading hashi; replacing them with different English words removes that clue. Translating a request before editing its politeness or force can likewise flatten distinctions that are difficult to recover. These are illustrative editorial examples, not reported experimental results from the papers.

Choosing a language for a real workflow

The following is an editorial proposal for applying the research. First define success: verification of a mathematical answer, preservation of conditions and proper names in a document, or understandable and faithful responses in a conversation. Then compare language arrangements against that criterion using equivalent input content.

If the model exposes control over the language of generated reasoning, record that as a separate condition. Do not infer internal computation language from a displayed explanation alone. Hold model version, tokenizer, generation limit, and sampling settings constant, and record input and output tokens, elapsed time, accepted results, and human repair time. Keep caching and tool-use conditions comparable too.

When Japanese expression itself is the subject, a sensible starting point is to retain the original and treat any translation as supplementary. For equations or formalized conditions, compare an English version after checking that meaning is preserved. A language arrangement that works well in mathematics should not be assumed to work equally well for internal documents or customer responses.

GoTo’s model is a useful reason to examine what regional adaptation actually improves rather than stopping at one multilingual score. Evaluation results can change with the same tokenizer; training can change reasoning-language behavior; and the appropriate language depends on the task. The practical objective is a configuration that completes the work accurately in a form its users can understand.