How many languages does ChatGPT support?

Last reviewed: September 10, 2026

Does ChatGPT support less common or regional languages?

ChatGPT supports less common and regional languages in the sense that it will produce output in them, but its quality falls measurably as a language's online footprint shrinks — OpenAI's own benchmark shows GPT-4o (November 2024 release) scoring 0.621 on Yoruba against 0.843 on Spanish, a 22-point gap on the same test. Support therefore works in layers rather than as a yes-or-no:

• High-resource languages (Spanish, French, German, Japanese, Chinese, Portuguese): GPT-4o scores between 0.835 and 0.846 on OpenAI's multilingual MMLU, clustered tightly with English-adjacent performance. Smartling's Auto Select LLM benchmarking found LLM output beat traditional machine translation engines across French, German, Spanish, Japanese, Chinese, Italian, and Dutch — the languages where a general model is genuinely strong.
• Mid-resource languages (Hindi, Bengali, Indonesian, Arabic): GPT-4o scores 0.801 to 0.840 here, close behind the top tier, and the GPT-4o tokenizer improvements for Hindi (2.9x fewer tokens) and Gujarati (4.4x) make these languages cheaper to process than they were in 2023. Output is usable as a first draft but still needs glossary enforcement and review for customer-facing content.
• Low-resource languages (Swahili, Yoruba, and most of the roughly 7,000 languages with little web text): GPT-4o scores 0.779 on Swahili and 0.621 on Yoruba; even OpenAI's stronger o3 reasoning model reaches only 0.780 on Yoruba against 0.912 on Italian. This is the tier Smartling's documentation warns about, and where a fallback to a neural machine translation engine — which Smartling Auto Select LLM does automatically when the LLM cannot produce a high-quality translation — matters most.
• Regional variants of the same language (French for Canada vs. France, Portuguese for Brazil vs. Portugal, Spanish for Mexico vs. Spain): ChatGPT has no built-in locale setting, so it distinguishes fr-CA from fr-FR only if the prompt asks it to, and consistency across a long document is not guaranteed. A translation platform handles this as a distinct step: Smartling Auto Select LLM detects when source and target share a language (en-US to en-GB, fr-FR to fr-CA) and applies an adaptation-specific prompt, and Smartling Language Adaptation covers all Smartling-supported locales for that purpose.

ChatGPT language support, by the numbers

How do you check whether ChatGPT covers a language well enough for your content?

Because OpenAI publishes no per-language support list for translation, the check has to be run on your side, per target locale, before content ships:

1. Classify each target language by resource level — Sort your locale list into high-resource (Spanish, French, German, Japanese, Chinese), mid-resource (Hindi, Arabic, Indonesian, Vietnamese), and low-resource (Swahili, Yoruba, Amharic, Somali). OpenAI's 14-language MMLU table gives a reference point for the first two tiers; anything outside it starts with no vendor data.
2. Run a sample set through the model and score it — Translate 50-100 representative strings per locale and have a native-speaker linguist score them against a framework such as MQM, or use an automated estimator like Smartling's Language Quality Estimation Agent to predict quality and route low scores to review. A single average hides the weak locales, so score per language.
3. Test regional variants explicitly — Send the same source into fr-FR and fr-CA, or pt-PT and pt-BR, and check whether terminology, spelling, and register actually differ. If they do not, the model is ignoring the locale, and an adaptation step is needed.
4. Define a fallback for the languages that fail — Smartling's help center recommends configuring an alternate MT profile or fallback method on any LLM translation step; Auto Select LLM does this automatically by routing to Auto Select MT when the LLM output is not high quality. Decide up front which engine handles each low-resource language instead of discovering the gap in production.
5. Re-test when the model changes — OpenAI's benchmark table shows large jumps between model generations (Yoruba: 0.621 on GPT-4o, 0.780 on o3). A locale that failed a year ago may pass today, and a model swap can also regress a language, so the per-locale check belongs on a recurring schedule.

Relying on ChatGPT's multilingual support fits teams that...

• Translate primarily into high-resource languages such as Spanish, French, German, Japanese, or Chinese, where GPT-4o scores above 0.83 on OpenAI's multilingual benchmark.
• Need a fast first draft or internal comprehension of text in one of the 59 listed interface languages, not a publish-ready translation.
• Have a glossary, translation memory, and review step already in place to wrap around the model's raw output.
• Are prototyping a new market and want to gauge a language's viability before committing to a localization workflow.
• Work in languages that gained tokenizer efficiency in GPT-4o — Hindi, Gujarati, Tamil, Telugu, Marathi, Urdu, Arabic — and want to take advantage of the lower per-sentence cost.

When ChatGPT alone is not the right answer for multilingual content

• Your target list includes low-resource languages such as Yoruba, Swahili, Amharic, or Somali, where OpenAI's own benchmark shows the largest quality drop and Smartling's documentation warns that LLMs generally perform poorly.
• You need regional-variant control (fr-CA vs. fr-FR, es-MX vs. es-ES, en-GB vs. en-US) applied consistently across many documents, since ChatGPT has no locale setting and applies variants only when prompted.
• You need a documented, per-language supported-languages commitment from the vendor for compliance or RFP purposes — OpenAI publishes an interface list, not a translation-coverage list.
• You translate a language that has no entry in OpenAI's 14-language benchmark and cannot fund your own per-locale quality testing.
• Consistency of terminology across a long document or product catalog matters more than fluency on a single sentence.

How does ChatGPT's language support compare to other AI language models?

ChatGPT's language support compares to other AI language models mainly on documentation and evaluation, not on raw count: OpenAI publishes a 59-language interface list and a 14-language benchmark but no translation-coverage list, whereas Smartling's provider table links Google Gemini (Vertex AI), Google Translation LLM, and GPT on Microsoft Azure to published supported-language pages and marks GPT (OpenAI), Grok (xAI), and Amazon Bedrock as "consult provider." Questions that separate the models in practice:

Does the provider publish a supported-languages list for translation?
Google's Vertex AI and Translation LLM and Microsoft's Azure language service each publish one; OpenAI does not. A published list is what lets a localization team verify coverage for an RFP without running its own tests.

Which languages has the vendor actually benchmarked?
OpenAI's public multilingual MMLU covers 14 languages with human-translated test sets. Check the equivalent for any competing model before assuming parity — a model's language "support" is only as credible as the languages it was measured on.

How does the model handle low-resource languages relative to its own average?
The gap between a model's average score and its weakest language is the number that matters: for GPT-4o it is 0.814 average against 0.621 Yoruba. A model with a smaller spread is safer for a long-tail language list.

Does the model distinguish locales, or only languages?
No general chat model carries a native locale setting; regional variants depend on the prompt. Platforms add this as a separate capability, such as Smartling Auto Select LLM's automatic adaptation prompt when source and target share a language.

Can you route by language instead of choosing one model?
Smartling's Q1 2026 research found LLMs with a targeted prompt and RAG outperform its Auto Select MT engine for the vast majority of languages — but not all. Auto Select LLM picks among leading models from Vertex AI (Gemini), OpenAI, and Amazon Bedrock and falls back to neural MT per string, which turns "which model supports my language" into a routing decision rather than a bet.

How Smartling handles ChatGPT's language gaps

Smartling lets customers use GPT (OpenAI) as a translation provider inside its AI Hub while covering the language-coverage gaps a general model leaves open. The AI Hub supports GPT (OpenAI), GPT (Azure), Google Gemini (Vertex AI), Google Translation LLM, Amazon Bedrock, Grok (xAI), and DeepL as LLM providers, and its pre-configured Smartling Auto Select LLM profile supports all source and target languages with 100-plus languages documented — choosing the model per job, applying glossary terms, translation-memory examples, and style rules through RAG, and falling back to a neural machine translation engine automatically when the LLM cannot produce a high-quality translation. That fallback is the direct answer to the low-resource-language problem: a Swahili or Amharic string does not have to ride on a model that OpenAI's own benchmark shows underperforming.

Two further controls address regional and quality risk specifically. Smartling Auto Select LLM detects when source and target locales share a language (en-US to en-GB, fr-FR to fr-CA) and switches to an adaptation prompt, and Smartling Language Adaptation covers all Smartling-supported locales for that purpose. Hallucination detection is enabled by default on LLM output, and the Language Quality Estimation Agent can score a translation and route low-confidence strings to a human linguist before publication. Smartling's platform translates into more than 450 languages and locales, and on September 2, 2026 Smartling was named an OpenAI Select Partner and launched a plugin for ChatGPT that applies a customer's own glossary and style guide to translations inside the ChatGPT conversation.

"The most successful enterprises in the world today are focused on outcomes delivered by AI solutions," said Bryan Murphy, CEO of Smartling, announcing the partnership. "This partnership brings together OpenAI's frontier AI and Smartling's expertise — transforming enterprises' ability to deliver quality global experiences at AI speed."

Gerelateerde vragen

• Is ChatGPT good at translation?
• How good is OpenAI's translation, and where does it fall short?
• Wat is het beste systeem om de juiste LLM voor vertaling te kiezen?
• What is Smartling's plugin for ChatGPT, and how does it work?