Seattle—September 4, 2025—Today, we are excited to announce a new Azure AI Translator API which provides you with flexibility to choose general neural machine translation (NMT), or a list of generative AI large language models (LLMs) at a request level. Using generative AI models, the new API offers new capabilities to produce tone translation, gender translation, and adaptive custom translation.
Generative AI LLMs produces more natural and fluent translations, and diverse translations based on specific instructions but often require more time to generate the translation. Translation with LLMs is more suited for localization, document translation, and workflows having human reviews. General NMT models produce more deterministic single translation at much lower latency better suited for real time translation scenarios.
Enterprises and organizations need machine translation for diverse use case scenarios. Each scenario has different business criticalities, functional and nonfunctional requirements. The new Azure AI Translator API allows customers to choose a model addressing each of those diverse needs at request level, providing control to them on quality, cost, latency, feature, and data residency.
Service capabilities (Preview):
- In a single request you can translate source text in different languages and using different models.
- During preview the API supports general NMT, gpt-4o and gpt-4o-mini models.
- Leverage generative AI LLMs to translate text in different tones – formal, informal, and neutral.
- Leverage generative AI LLMs to translate text in different gender – male, female, and neutral.
- Provide up to five reference translations for generative AI LLMs to treat them as a few shot examples and produce translations with similar style.
- Languages evaluated and supported for generative AI LLM translations are listed here.
Pre-requisites:
Customers can use either ‘Azure AI Translator’ or ‘Azure AI Foundry’ resource to produce translation using general NMT models, like today.
However, to leverage new capabilities and produce translation using LLMs, customers need an ‘Azure AI Foundry’ resource. During preview, customers need to deploy their own instance of the LLM in the same Azure AI Foundry resource. Currently the API supports gpt-4o and gpt-4o-mini models.
Authentication:
The API supports authentication using both resource key and Entra ID.
API Requests:
In the new Translator API, you will specify translation parameters in the request body instead of passing them as query parameters.
Here is the curl command on using Azure AI Translator global endpoint (recommended).
curl -X POST "https://api.cognitive.microsofttranslator.com/translate?api-version=2025-05-01-preview" \
-H "Ocp-Apim-Subscription-Key:<your-key>" \
-H "Ocp-Apim-Subscription-Region:<your-resource-region>" \
-H "Content-Type: application/json" \
-d "<your-request-body>"
Here is the curl command on using custom endpoint.
curl -X POST "https://<your-resource-name>.cognitiveservices.azure.com/translator/text/translate?api-version=2025-05-01-preview" \
-H "Ocp-Apim-Subscription-Key:<your-key>" \
-H "Ocp-Apim-Subscription-Region:<your-resource-region>" \
-H "Content-Type: application/json" \
-d "<your-request-body>"
Below are sample request-body for few use cases.
Translate text using general NMT
[
{
"Text": "Doctor is available next Monday. Do you want to schedule an appointment?",
"language": "en",
"targets": [
{
"language": "es"
}
]
}
]
Translate text using generative AI LLM
A gpt-4o-mini model instance with deployed name “contoso-gpt-4o-mini” is used. On not specifying the source language the system will detect it automatically and produce translation.
[
{
"Text": "Doctor is available next Monday. Do you want to schedule an appointment?",
"targets": [
{
"language": "es",
"deploymentName": "contoso-gpt-4o-mini"
}
]
}
]
Translate text using both general NMT and generative AI LLM in a single request
In this sample the source text is translated into Spanish language using gpt-4o model and into German language using general NMT model.
[
{
"Text": "Doctor is available next Monday. Do you want to schedule an appointment?",
"targets": [
{
"language": "es",
"deploymentName": "contoso-gpt-4o "
},
{
"language": "de",
}
]
}
]
Gender specific translation
Leverage generative AI LLMs to produce gender specific translation.
[
{
"Text": " Doctor arrives at 10:00AM today.",
"language": "en",
"targets": [
{
"language": "es",
"deploymentName": "contoso-gpt-4o-mini",
"gender": "female",
},
{
"language": "es",
"deploymentName": "contoso-gpt-4o-mini",
"gender": "male",
}
]
}
]
Considering not all languages have gender specific terms, and some languages don’t have gender-neutral terms, the table below guides you on expected gender in the translation.
Source text |
Target gender |
Expected gender in translation |
Neutral |
Not indicated |
Neutral or Random: if target language doesn't have a gender-neutral term. |
Female |
Not indicated |
Female or Neutral: if target language is gender neutral. |
Male |
Not indicated |
Male or Neutral: if target language is gender neutral. |
Neutral |
Female |
Female or Neutral: if target language is gender neutral. |
Female |
Female | |
Male |
Female | |
Neutral |
Male |
Male or Neutral: if target language is gender neutral. |
Female |
Male | |
Male |
Male |
Customize translation with references
Leverage generative AI LLMs to produce custom translation using translated reference sentences for similar terms or style.
[
{
"Text": " Doctor arrives at 10:00AM today.",
"language": "en",
"targets": [
{
"language": "es",
"deploymentName": "contoso-gpt-4o-mini",
"referenceTextPairs": [
{
"source": "text_in_en",
"target": " text_in_es"
},
{
"source": " text_in_en",
"target": " text_in_es"
}
}
]
}
]
Service limits:
The service limits differ for translation using general NMT models and generative AI LLMs.
Operation |
Maximum Number of Array Elements |
Maximum Size of Array Element |
Generative AI LLM: Maximum Number of Array Elements |
Generative AI LLM: Maximum Size of Array Element |
Translate |
1,000 |
50,000 |
50 |
5,000 |
Latency of translation using generative AI LLMs varies based on the capacity provisioned by customers during model deployment.
Pricing
Translations using general NMT models or generative LLMs are charged as per respective rate card.
- Translations using general NMT models, which are default, are charged based on characters in source text as per price Azure AI Translator pricing.
- Translations using generative AI LLMs are charged based on input and output tokens processed as per Azure Open AI pricing.
Other changes in the enhanced Translator API:
The new Translator API retains translation and transliteration functionality. It deprecates legacy non-core translation features. Please see more details here.
Resources:
Azure AI Translator | Microsoft Azure
Azure AI Translator documentation