DeepL vs Google Translate vs LLMs: What 100,000 Words Cost and What WMT Proves

Translating a 100,000-word documentation set into 5 languages costs about $50 on Google's standard neural engine and about $15 on Claude Sonnet 5.5, on November 2026 rate cards. The deepl vs google translate question is no longer mainly about price, because a mid-tier LLM now undercuts both. It is about quality, and the only peer-reviewed evidence on quality cannot tell you which of DeepL or Google is better.

The short answer. For a fixed documentation job, a mid-tier LLM such as Claude Sonnet 5.5 or GPT-6.1 Sol costs a fraction of DeepL or Google Cloud Translation at list price. Peer-reviewed WMT human evaluations rank general LLMs at or near the top. They anonymise DeepL and Google, so any claim that one beats the other rests on vendor tests.

I have written this for the person who owns localisation delivery, the head of documentation or operations who has to pick an engine and defend the line item. That reader converts on a priced workload and a checklist, not on a leaderboard.

The 100,000-word job, priced on every rate card

A fixed workload makes rate cards comparable. Ours is a 100,000-word English documentation set, translated into German, French, Spanish, Japanese and Brazilian Portuguese. That is a realistic first localisation wave for a B2B product.

The prices below are list rates I retrieved on 3 November 2026 from each vendor's own pricing page. Everything else in the table is my arithmetic, and the assumptions behind it are stated next.

The assumptions, stated so you can change them

Character-metered engines need a character count. I assume 6 characters per English word, counting spaces and punctuation, which gives 600,000 characters per target language and 3,000,000 for all five. Every half-character you move that assumption shifts the character-metered rows by about 8%.

Token-metered engines need a token count. OpenAI's own help article says 1 token is approximately three-quarters of a word. That makes 133,333 English input tokens. I add 20% for instructions, glossary and context, giving 160,000 input tokens per language.

Output is the uncertain part. I assume each target language needs 1.5 times as many tokens as the English source, which is my assumption rather than a published figure. For Claude I add a further 30%, because Anthropic says its newer tokenizer produces approximately 30% more tokens for the same text.

Token and character volumes are our assumptions, not vendor figures. The DeepL Growth fee buys 12 million characters a year, enough for 4 runs of this job, so $78 is the effective per-run cost if you use the full allowance. Gemini 3.8 Flash rises to $1.50 and $7.50 per million tokens on 1 January 2027, which doubles that row to $8.70.

Two things stand out. First, the cheapest serious LLM option costs less than a quarter of Google NMT. Second, DeepL is the most expensive way to run a single job, but only because its Growth plan is an annual commitment. If you translate 12 million characters a year, its effective rate drops to $26 per million, close to Google's $20.

My position: at these prices, the per-unit translation fee should no longer be the criterion you use to pick an engine. The spread between the cheapest and most expensive realistic option is a few hundred dollars. A single round of human review on 100,000 words will cost more than every row in that table combined.

How DeepL and Google meter characters differently

Character pricing looks simple and hides two choices that change your bill: which characters count, and whether output counts. The two vendors made different choices on both.

DeepL bills source characters only

DeepL's developer documentation says it bills by the number of characters in your source text, measured in Unicode code points. Text inside HTML or XML markup is not billed when tag handling is on. That matters for documentation, where markup can be a large share of every page.

Two DeepL rules can surprise a docs team. Characters still count when the source and target languages are the same. Document uploads such as .docx and .pdf are billed at a minimum of 50,000 characters per document, so sending 300 short help articles as separate files is an expensive mistake.

The current DeepL API plans are Developer, a free trial with a one-time credit of 1 million characters, and Growth. Growth costs $26 a month billed annually, includes 12 million characters a year, caps usage at 50 million characters a month and charges $27.50 per million extra characters. Enterprise pricing is custom.

Google bills whitespace, markup and, for LLM tiers, output

Google's rate card is explicit that you are charged for every character in a request, including whitespace and untranslated characters. Its own example says a short Japanese greeting wrapped in paragraph tags counts as 12 characters. If your docs pipeline sends raw HTML, you pay for the tags.

The standard NMT tier charges only on input. The two LLM tiers charge on input and output separately: Translation LLM at $10 per million characters each way, and Adaptive Translation at $25 each way. Google describes Translation LLM as cost equivalent with NMT, which is only true if output length matches input length.

I think the Translation LLM tier is the most underrated option on Google's card. It brings a language model to a per-character price with no token estimation, which removes the forecasting problem the next section describes.

How LLM translation is billed per token

LLM APIs from Anthropic, OpenAI and Google price per million tokens, with output typically 5 times the input rate. A token is a chunk of text whose size depends on the model's tokenizer and on the language. That is why LLM translation cost is a forecast, while DeepL and Google NMT cost is a count.

The list rates are close at the mid tier. Claude Sonnet 5.5 and GPT-6.1 Sol both charge $2 input and $10 output per million tokens. Gemini 3.8 Flash charges $0.75 and $3.75 until the end of 2026. All three vendors offer a 50% discount through their batch APIs, which suits a documentation backlog that does not need answers in seconds.

The batch discount is the single largest lever in this section. It halves our Sonnet 5.5 figure to $7.54 and our GPT-6.1 Sol figure to $5.80. The trade-offs are covered in the comparison of prompt caching and batch discount rates, and a wider rate table sits in the inference provider rate card.

The tokenizer tax on non-English output

OpenAI's rule of thumb is written for English. Its own help article warns that other languages can have different relationships between characters, words and tokens. In practice that means your German or Japanese output may need noticeably more tokens than the English source, and output is the expensive side of the bill.

I used a 1.5 times multiplier. To stress it, I reran Sonnet 5.5 at 3 times, which is a deliberately pessimistic case. The job rises from $15.08 to $28.08. That is still below Google NMT at $50, so the conclusion survives the assumption I trust least.

Reasoning tokens can multiply the bill

The bigger risk is a model that thinks before it answers. The WMT25 organisers did not set a reasoning budget for Gemini 2.5 Pro, and its translation of the test set cost $250.8 against $31.7 for GPT-4.1. The preliminary WMT25 ranking paper says the missing budget increased output token count 6.6 times.

The lesson for your pipeline is mechanical. Turn reasoning off, or cap it, for translation calls, and log output tokens per request. An uncapped reasoning model can turn a $15 job into a three-figure one with no visible change in quality. Why identical tasks produce very different token bills is examined in the analysis of agent token cost variance.

What WMT human evaluation shows about LLM translation quality

The Conference on Machine Translation, WMT, runs the field's longest-standing shared task. Its findings papers are peer reviewed and its rankings come from professional human annotators, not automatic metrics. That makes it the right source for llm translation quality claims, and the wrong source for a few claims people attribute to it.

Both recent editions used Error Span Annotation, or ESA. Annotators mark error spans in each translation and give a 0 to 100 score. Systems are then grouped into statistical clusters, and WMT counts a system as winning a language pair when it sits in the top cluster.

WMT24: Claude-3.5-Sonnet won 9 of 11 pairs

WMT24 covered 11 language pairs and added translations from 8 LLMs and 4 online translation providers to the participant systems. The findings paper says the best performing system in general is Claude-3.5-Sonnet, winning in 9 language pairs. It also notes Claude outperformed GPT-4, which won 5, and which was much more expensive.

The cost table is worth reading. Translating the WMT24 test set cost $31.9 with Claude-3.5 and $240.4 with GPT-4. The paper itself notes that more recent GPT models are significantly cheaper, which is why I use current rate cards above rather than these figures.

The pair-level scores show how close the top is. In English to Japanese, the anonymised ONLINE-B scored 91.1, ahead of GPT-4 and Claude-3.5 at 90.8. In English to Spanish, GPT-4 scored 91.9 while ONLINE-B ranked 11th at 85.6.

WMT25: Gemini 2.5 Pro in the top cluster for 14 of 16

WMT25 evaluated 60 systems, 36 from participants and 24 collected from LLMs and popular online translation providers. Its human evaluation covered 16 language pairs on harder, document-level test sets. The findings paper says Gemini 2.5 Pro places in the top cluster for 14 of the 16 evaluated language pairs and is on par with or surpasses human translation in 10 of those pairs.

The same paper is blunt about what remains unsolved. Human references were in the winning cluster for only 6 of 15 pairs where they existed. It also reports that even the top-performing system often fails to score above 90, and that the speech domain was hardest while literary text was easiest.

The contrarian reading I hold: WMT25 is better news for buyers than its headline suggests. If general LLMs match professional human references in most pairs, the question for your documentation is not whether machines can translate it. It is how you find the segments where they fail, which is a review design problem.

DeepL vs Google Translate: the comparison WMT cannot make

Search for deepl vs google translate and you will find posts citing WMT as proof of one or the other. That citation does not hold. The WMT24 paper says online system translations were obtained by the organisers rather than submitted, and are therefore anonymised. WMT25 does the same, labelling its three public MT services ONLINE-B, G and W.

Neither paper names which label belongs to which company. Any article that tells you ONLINE-B is DeepL or Google is guessing, and I am not going to repeat the guess.

The only numbered head-to-head is DeepL's own. In July 2024, DeepL announced that language experts preferred its next-generation model 1.3 times more often than Google Translate and 1.7 times more often than ChatGPT-4. It also said Google needed 2 times more edits to reach the same quality.

The DeepL claim has a real limitation that its announcement states nowhere. It publishes no evaluator count, no segment count, no language pairs for the competitor comparison, no test dates and no competitor versions. It is also more than two years old, and the competitor models it beat have since been replaced.

That does not make DeepL worse. It means the honest answer to "is DeepL better than Google" is that no independent, published study you can open today settles it. Run a pilot on your own content. The method for that is in the guide to building an LLM evaluation suite.

Where this argument is weakest

The table at the top of this post is precise and incomplete. Here is what it leaves out.

The review cost is missing, and it is the largest number

I priced machine output only. A documentation team that ships translated docs without review accepts the error rate WMT25 describes, where even the best system often scores below 90. Most teams will pay translators to post-edit, and that labour is priced per word or per hour by your vendor.

I have not published a post-editing rate because I could not find one with a stated sample and methodology. Get a quote from your language service provider. If an engine cuts edits per segment, it can be worth its higher machine price many times over, which is the real argument DeepL's edit-count claim is making.

Benchmarks are not your documentation

WMT tests news, social, speech and literary text. Product documentation is full of UI strings, code blocks, product names and terminology that must stay fixed. DeepL glossaries, Google custom models and LLM system prompts handle terminology in different ways, and none of the WMT results measure that.

There is also a case where the per-character vendors win outright. They return deterministic, auditable billing with no token forecast, no reasoning budget and no prompt to maintain. For a team without an engineer to own a prompt, that predictability is worth more than the savings in our table. Why rankings shift with test design is laid out in the explainer on why LLM leaderboards disagree.

Choosing an engine for B2B documentation

The right engine depends on who owns the pipeline and how much you translate, more than on which model leads a leaderboard. This checklist reflects the evidence above.

If you route between models, keep the router honest about quality as well as price. The failure mode is covered in the analysis of the model routing quality gap. If you use an LLM to score translations, read the review of LLM judge agreement and bias studies first, because WMT25's own automatic ranking used GPT-4.1 as a judge while also ranking it.

Frequently asked questions

Where to start

Pick 2,000 words from your real documentation, including UI strings and a code sample. Run them through Google Translation LLM, DeepL and one mid-tier LLM with your glossary in the prompt, then ask one reviewer per language to count edits per segment, blind to the engine.

That count, multiplied by your post-editing rate, is the number that should decide your engine. The machine price in the table above will not move your budget much. The reviewer hours will.

Related reading