Translator 2026-06-06 LLM translation: are source characters billed as well as tokens?

Question

Translator 2026-06-06 LLM translation: are source characters billed as well as tokens?

Jade Wilson 0

We're looking at moving from NMT to LLM translation with the Text Translation API 2026-06-06, and we can't tell from the docs how LLM requests are billed.

The overview page says:

▎ - By default, translations using general NMT (Neural Machine Translation) models are billed according to the number of characters in the source text.

▎ - Translations using generative AI LLMs are charged according to the number of input and output tokens processed.

But the translate method examples show LLM responses reporting both characters and tokens as charged. For the GPT-5.1 example:

"sourceCharactersCharged": 72

"sourceTokensCharged": 26

"targetTokensCharged": 16

Our questions:

When a request uses an LLM deployment (deploymentName set to a GPT deployment), is the S1 per-character Translator charge also applied to the source characters, on top of the Azure OpenAI token charge for that deployment?
If not, what does sourceCharactersCharged mean on an LLM response?
Are the instruction tokens Translator adds (instructionTokens in the response) billed as input tokens at the deployment's rate?

Also, is there a published list of models supported for LLM translation in 2026-06-06? Deployments of gpt-4.1-mini and gpt-5-mini return "Deployment ... is currently not supported for 2026-06-06", and the docs only name GPT-5.1, GPT-5.4, GPT-4o and GPT-4o-mini.

0 comments

1 answer

Your answer

Answer 1

Hello Jade Wilson

Based on the current Azure Translator 2026-06-06 documentation, LLM translation is intended to use token-based billing rather than an additional S1 per-character charge.

For your specific questions:

1)Are source characters charged in addition to GPT tokens?

Translator pricing documentation distinguishes NMT (character-based) from LLM translation (input/output token-based).

The presence of sourceCharactersCharged in an LLM response does not establish that a second charge is applied.

2)What does sourceCharactersCharged mean?

It reports source-character consumption, but the current documentation does not clearly specify whether this field contributes to a separate billing meter for LLM requests.

You can verify this through Azure Cost Management → Cost analysis, grouping by meter and resource, whether the request generates both Translator character charges and Azure OpenAI token charges.

3)Are instructionTokens billed?

Yes, the documented example strongly indicates they are included in billable input tokens.

For GPT-5.1, for example:

instructionTokens = 12
sourceTokens = 14
sourceTokensCharged = 26
targetTokensCharged = 16

Especially for short translations where instruction tokens can represent a significant portion of the request.

Official Translate API examples

4)Which LLM models are supported?

Microsoft documents GPT-5.1 and GPT-5.4 as examples of supported LLM deployments, but I could not find a complete, authoritative compatibility list specifically for Translator API 2026-06-06.

The errors from gpt-4.1-mini and gpt-5-mini indicate those deployments are not currently accepted by this Translator integration, even though the models may work independently through Azure OpenAI.

Before migrating production workloads, run a small isolated LLM translation test and compare the Azure OpenAI token metrics with the actual billing meters in Cost Management.

If both character and token charges appear, open a Microsoft Support case with the request ID and billing details.

Official documentation: