Guide

Choosing a translation API in 2026

How to read this

All these services bill per character of source text on their standard tariffs, and they bill the same characters again for each target language. Translating 1,000 characters into three languages bills 3,000. Google's Translation LLM also bills output characters, and Azure's GPT-based option bills tokens.

We read vendor prices from each vendor's pricing page or documentation in November 2026. We give prices in US dollars unless marked otherwise. We name annual commitments where they apply. For the LLM figures, we converted what our benchmark runs cost to the same basis.

Pricing

Azure AI Translator charges the least of the large services, at $10 per million characters, and DeepL the most, at $27.50 beyond its included amount.

Service	Price per million characters	Free allowance	Notes
Azure AI Translator, S1	$10	2 million a month on the separate F0 tier	Lower rates with a monthly commitment, $8.22 down to $5.50
Google Cloud Translation, Translation LLM	$10 input plus $10 output	none listed	Output characters are billed too
Amazon Translate	$15	2 million a month for 12 months	Starts with the first request
Google Cloud Translation, NMT	$20	500,000 a month	Applied as a $10 monthly credit
Lara Translate, Standard model on Pro	$24.99 (€20)	10,000 a month	Needs an active subscription
DeepL API, Growth	$27.50 beyond the included amount	Developer plan, 1 million, once	$26 a month billed annually, 12 million a year included
Five LLM setups, from our runs	$0.10 to $0.21	depends on the provider	Measured cost, see below

Read in November 2026. Service prices are usage rates per million source characters per target language.

Azure's pricing page has the lowest rate of the large services, at $10 per million on its S1 tier. Its separate free tier, F0, gives 2 million characters a month, shared with custom model training, and does not expire. Commitment tiers start at $2,055 a month for 250 million characters, with $8.22 per million beyond that. At 4 billion characters a month, the rate beyond the commitment is $5.50.

Amazon Translate charges $15 per million characters. Its free tier gives 2 million characters a month for 12 months from your first request. Its custom model option, Active Custom Translation, costs $60 per million.

Google's pricing page lists two ways to translate text. The NMT model, the long-standing default, costs $20 per million after the first 500,000 characters each month. Google applies that free amount as a $10 credit, shared by the Basic and Advanced editions. Above a billion characters a month, you agree NMT pricing with Google's sales team. The Translation LLM costs $10 per million input characters and $10 per million output characters. Google's page calls that "cost equivalent with NMT".

DeepL changed its plans. According to DeepL's support page on API plans, you can no longer buy the old API Free and API Pro plans. Existing Free accounts keep 500,000 characters a month, which is what our test key reported. New accounts start on the Developer plan, which gives 1 million characters once and never resets. The paid Growth plan costs $26 a month, billed annually. It includes 12 million characters a year and charges $27.50 per million beyond that.

Lara Translate, from the company behind ModernMT, charges $24.99 per million for its Standard model on the Pro plan. Its other models cost more, and Lara mentions a platform fee without publishing the amount. Annual volume plans bring the Standard rate down to €5 to €12 per million.

ModernMT's pricing page now says ModernMT shuts down at the end of 2026. Existing keys work until the end of December 2026.

What an LLM costs on the same basis

On the same per-character basis, the LLM setups we ran cost $0.10 to $0.21 per million characters, against $20 for Google NMT and $27.50 for DeepL. LLMs bill per token, and the token count includes your prompt. For this comparison we ran our benchmark on FLORES-200. It is a set of sentences with professional translations in every language. Each setup got the same 700 source sentences, 2,000 translations in all across 20 language pairs. Each LLM call included our full prompt and asked for every target language of that sentence. Google billed the same run as 266,661 characters, so we divided the bill for each setup by that number.

Setup	Cost of the run	Per million characters
DeepSeek V4.1 Flash, Morph host	$0.025	$0.10
Gemini 2.5 Flash-Lite	$0.048	$0.18
DeepSeek V4.1 Flash, Wafer host	$0.052	$0.20
GPT-6 Luna	$0.056	$0.21
Gemma 4 31B	$0.057	$0.21
Google Cloud Translation, NMT	$5.33	$20.00
DeepL	$7.33	$27.50

FLORES-200, November 2026. Costs at list usage rates, without free allowances or plan fees.

On a log scale, where each step is ten times the last, the two groups sit far apart.

The cheapest setup therefore costs about 200 times less than Google's NMT and about 290 times less than DeepL. Not every LLM costs that little. In the same run, gpt-oss-120b with its thinking on came to $0.82 per million characters. The two DeepSeek rows are the same model. The Wafer run cost about twice the Morph run. We compare those hosts in DeepSeek API pricing across hosts.

Live captions cost an LLM more per character. The LLM translates each sentence while the speaker is still speaking. In our live test of rooms translated into three to five languages, DeepSeek V4.1 Flash on Wafer cost over four times as much per character as it did on FLORES. We counted characters the way Google bills them. We sent the services only finished sentences, so they billed each character once. Even so, a room cost about 23 times as much on Google and 31 times as much on DeepL. Room cost here means what translation costs for one hour of a room, one person speaking with captions in three to five languages, and the results table below gives it as a multiple of DeepSeek V4.1 Flash on Wafer.

Free tiers for live captions

Free allowances go quickly with live captions, because every target language counts the characters again. A room in four languages uses an allowance four times as fast as captions in one. Azure and Amazon each give 2 million characters a month, four times Google's 500,000. Amazon's amount lasts only for the first year. DeepL's Developer plan gives 1 million once. Lara gives 10,000 a month, a fiftieth of Google's.

Set your own monthly volume and number of languages to see which free tiers last.

Features

DeepL and Lara take context with each request and Google's NMT model takes none, while no service we read guarantees that its glossary appears in every output. With an LLM, the glossary, context and formality all go in the prompt.

Service	Languages	Glossary	Context	Formality	LLM option
Azure AI Translator	over 100, fewer on the LLM option	dynamic dictionary, custom models	reference examples on the LLM option	on the LLM option	GPT-5.1, billed per token
Amazon Translate	75	custom terminology	not listed	on supported languages	not listed
Google Cloud Translation	189 advertised, fewer on the LLM	yes	not on NMT	not listed	Translation LLM
DeepL API	126 entries, counting variants	up to 2,000 on Growth	yes, not billed	on supported languages	quality or speed setting
Lara Translate	over 200	glossaries, translation memory	context, instructions, style guides	not listed	is an LLM
An LLM API	depends on the model	in the prompt	in the prompt	in the prompt	is an LLM

Read in November 2026 from each vendor's documentation. "Not listed" means the pages we read did not mention it.

DeepL's translate endpoint takes a context field for text that guides the translation. DeepL does not translate or bill for that text. We sent the previous sentence there, because a caption is often half a sentence. We did not measure how much it helped. Google's NMT model takes no context.

Azure's GPT-based option is part of its generally available 2026-06-06 API. It accepts up to five example translations, plus tone and gender settings. It covers fewer languages than Azure's standard model. Azure bills it by tokens at Azure OpenAI rates, not per character.

On glossaries, Amazon's custom terminology docs say Amazon applies a term in context, so it may not appear every time. Azure's dynamic dictionary needs English on one side. Microsoft says it is safe only for compound nouns like product names. With an LLM, you put the glossary in the prompt as a list and check the output yourself.

DeepL's model_type setting asks for lower latency, the default, or higher quality, according to DeepL's model types page. DeepL treats it as a goal and chooses the underlying model itself. We asked for quality.

Google's product page advertises 189 languages. Its Translation LLM supports a shorter list, so check the per-model language table before you choose it. DeepL's supported languages page shows 126 entries. DeepL counts variants like Brazilian Portuguese as separate entries. It also includes languages DeepL translates into but not from.

Document pricing differs

You pay different prices if you translate files instead of text. Azure charges $15 per million characters for documents. Amazon charges $15 per million for batch documents, the same price as text, but $30 for real-time Word files. Google charges $0.08 per page for Word, PowerPoint and PDF files on NMT. Google bills Excel files per character, like text. Google gives no free monthly credit for formatted documents. Lara bills each paid document for at least 20,000 characters.

What happens to your text

Azure's standard model, Google and DeepL's paid Growth plan keep your text private with nothing to set. Amazon, Lara, Azure's GPT-based option and an LLM through OpenRouter may keep it unless a setting or an application limits it, and on DeepL's free plans only an upgrade to Growth helps. The chart groups the services by what you have to do.

Service	Is your text kept?	Used to train models?	How to keep it private
Azure AI Translator, standard	not written to persistent storage	no	nothing to set
Azure, GPT-based option	may be kept for abuse review	not without permission	eligible customers can apply to turn that off
Google Cloud Translation	held briefly in memory	no	nothing to set
DeepL API, Growth (paid)	only while translating, plus rare debugging copies up to 72 hours	not outside your account	comes with the paid plan
DeepL API, Developer (free)	DeepL may keep it permanently	not stated	upgrade to Growth
DeepL API Free, older plan	may be kept for a limited period	may be	upgrade to Growth; no switch stated
Amazon Translate	may be kept	may be used to improve Amazon's AI	an opt-out policy stops the training use
Lara Translate	may be kept by default	may be, by default	Incognito mode on each request
An LLM through OpenRouter	depends on the provider	depends on the provider	zdr: true, plus "ZDR only" in account settings

From each vendor's privacy and data pages, read in November 2026, for ordinary text requests. Saved glossaries, custom models and document jobs have their own rules.

DeepL's free and paid plans have different terms. DeepL's terms say it "reserves the right to perpetually store" what you send through the free Developer plan. They also forbid sending personal data through it.

On the paid Growth plan, DeepL keeps ordinary requests only as long as needed to translate them. In rare cases it stores an encrypted copy for up to 72 hours to debug an error. Text you ask DeepL to save is deleted 90 days after your agreement ends. DeepL's privacy policy says content from its free service, which the older API Free plan falls under, is processed "for a limited period of time to train and improve" its models.

Google and Azure need no privacy setting for ordinary text requests. Google's data usage page says text "is held briefly in-memory in order to perform the translation". It also says Google does not use that text to train or improve its translation features.

Microsoft's No Trace page says text sent to Translator is "not written to persistent storage" and not used for training. That covers the standard and custom models. The GPT-based option runs on Azure's AI platform instead, where prompts can be stored for human abuse review. Eligible customers can apply to have that turned off, and Microsoft has to approve it.

Amazon Translate's FAQ says it "may store and use text inputs" to improve Amazon Translate and other Amazon AI. Some of that text may be stored in another AWS region. An AI services opt-out policy for your AWS organization stops that improvement use and deletes the copies kept for it. It does not cover content Amazon needs to run the service.

Lara's default Learning Mode allows it to keep text and use it for training. Its translate endpoint has a no_trace setting, off by default. Turning it on enables Incognito mode, where Lara discards the text after translation.

With an LLM, it depends on who runs the model. OpenRouter's data collection page says it does not store prompts or responses by default. Turning on its logging or its product improvement option lets it store them. Content flagged as possible child abuse material may also be kept for review, reporting and legal preservation.

The zdr setting on a request follows a choice in your OpenRouter account. OpenRouter's zero data retention page says self-serve accounts that have not chosen also allow providers under "Safety Retention", which may keep prompts to detect abuse. Choosing "ZDR only" in the account settings excludes those providers.

What we measured

On live speech, DeepL scored 96.2, below GPT-6 Luna and DeepSeek V4.1 Flash, and Google NMT scored lowest of these five setups at 92.3, though all five were within 0.006 COMET on standard sentences. We tested DeepL and Google NMT the way a live caption service could use them. We sent one request per target language for each finished sentence, in parallel. DeepL got the sentence before as context. Google NMT cannot take context. Two AI reviewers, Claude Opus 5.5 and GPT-6.1 Sol, scored every setup's translations of the same 220 passages from seven recorded talks. The full benchmark is in our LLM translation test.

Setup	COMET	Live quality	First words	Locked	Room cost
GPT-6 Luna	0.897	98.1	3.4 s	2.8 s	1.0x
DeepSeek V4.1 Flash, Wafer	0.896	97.2	1.5 s	1.4 s	1.0x
DeepL	0.897	96.2	6.1 s	2.1 s	31x
Gemini 2.5 Flash-Lite	0.892	94.1	1.3 s	1.2 s	1.2x
Google Cloud Translation, NMT	0.898	92.3	4.0 s	0.75 s	23x

Measured in November 2026. COMET is the FLORES score. We scored live quality from 0 to 100 for all five setups in one round. First words and locked are medians: the time from a sentence starting until any translation appears, and from a sentence ending until its final translation. Room cost covers translation only at list usage rates, as a multiple of DeepSeek V4.1 Flash on Wafer.

On FLORES, these five setups are too close to rank. On live speech, the spread is six points. DeepL scored 0.9 below DeepSeek and 3.9 above Google.

Their best and worst pairs differed. DeepL beat DeepSeek on Chinese into English and into Korean. Into Korean, DeepL scored 97.2 and DeepSeek scored 95.6. DeepL did not beat DeepSeek on Chinese into Vietnamese. DeepL was weakest from Spanish, at 92.0 into Portuguese against DeepSeek's 98.1. Google scored 98.3 from English into Chinese and 82.5 from Chinese into Korean. We did not test whether giving Google context would close part of its gap.

Google answered a finished sentence in 0.18 seconds at the median, DeepL answered in 0.46 seconds, and the LLMs in this table answered in 0.47 to 1.4 seconds.

We sent the services finished sentences only, so their captions waited for each sentence to end. The first words of a sentence showed up 4.0 seconds after it started on Google and 6.1 seconds on DeepL. The LLMs also translated the unfinished sentence and showed them in 1.3 to 3.4 seconds. DeepL's free plan also slowed some requests. Of 567 calls, 61 took over 3 seconds.

Which to pick

Use an LLM if cost per character matters and you can write and test a prompt. In our live test, a room on DeepSeek V4.1 Flash cost about a 23rd of Google's cost and a 31st of DeepL's cost. The LLM tied both services on standard sentences and scored above both on live speech. You maintain the prompt. Your code parses the output and handles retries.

Use a service if you want built-in terminology controls and document handling. Check language coverage and limits first. If your text is private, Google and Azure's standard model need no privacy setting for ordinary text requests. DeepL's free plans and Amazon's defaults allow your text to be kept beyond the translation itself. In our test, DeepL scored best of the services. Azure has the lowest rate of the large services, at $10 per million, and the largest free tier that does not expire.

If most of your source text is Chinese, test DeepL against your LLM before choosing. DeepL beat DeepSeek on two of the three Chinese pairs we tested.

If you need a single finished sentence answered as fast as possible, Google NMT was the fastest call we measured, at 0.18 seconds.