People use ChatGPT in Urdu, Arabic, Chinese and dozens of European languages every day, and many notice that the quality of the answers is not the same across them. The short explanation is that the model performs best in English, because that is the language it has seen most. The longer explanation says a lot about how these systems are built and what it would cost to build one that treats every language equally.
English dominates the training data
Large language models learn from vast collections of text scraped from the internet, books and other sources. English makes up a disproportionate share of that material, so a model like ChatGPT receives far more examples, more variety and higher-quality text in English than in any other language. When a user asks a question in English, the answer tends to be more accurate, more fluent and more detailed as a result.
The imbalance is compounded by how models break text into pieces, called tokens. Older tokenizers were tuned for English and split languages written in other scripts into many more fragments, which made processing those languages slower, more expensive and often less accurate.
Newer models have narrowed the gap
OpenAI has made multilingual performance an explicit target. When it launched GPT-4o in May 2024, the company said the model used a new tokenizer that represents many non-English languages far more compactly; in its launch materials OpenAI listed Urdu among the languages that needed roughly 2.5 times fewer tokens than before. Fewer tokens means more of a conversation fits in the model’s working memory and the model handles the language more efficiently. GPT-4.5, released in February 2025, continued that trend, with OpenAI reporting gains across a range of languages.
In practice, users working in Spanish, French, German and Urdu report that recent models no longer simply translate from English in their heads. They can write original text, follow idioms and hold a natural conversation in the target language, and they are better at picking up cultural context and register.
Where performance still lags
The improvement is uneven. For languages with little online text, including many African languages and regional dialects, answers can still be thin, generic or wrong. The problem is a data problem as much as a modeling one: if there is not much written material in a language, there is not much for the model to learn from.
Even in a comparatively well-served language like Urdu, English usually remains the better choice for specialized, technical or obscure topics. Documentation, research papers and technical discussion are overwhelmingly published in English, so the model simply knows more when asked in that language.
Could an Urdu-first model match ChatGPT?
Building a model that performs in Urdu the way ChatGPT performs in English is possible in principle, but the obstacles are practical. It would require a very large and diverse corpus of Urdu text covering idiom, sentence structure, regional variation and every subject area users might ask about. Much of that material does not exist in digital form, and collecting, cleaning and licensing it would be a major undertaking in its own right.
Training is the other cost. Frontier models are trained on thousands of specialized chips over weeks or months, and the compute alone runs to tens or hundreds of millions of dollars for the largest systems. A smaller, language-focused model would be cheaper, but still a multimillion-dollar project once data collection, training, evaluation and ongoing updates are counted. For most organizations the more realistic route is to fine-tune an existing open model on Urdu data rather than train one from scratch.
What users can do now
Model choice matters: among OpenAI’s models in mid-2025, GPT-4o is the strongest everyday option for Urdu. The quality of the prompt matters too. Clear, specific and well-structured questions produce better answers in every language, and switching to English for highly technical topics remains a sensible workaround while the data gap persists.
