What if the price tag on your translation bill has almost nothing to do with the price you saw on the vendor’s website? That gap trips up more teams than almost any other part of buying translation software. A marketing lead compares four pricing pages, picks the cheapest headline number, and then watches the invoice land three times higher once glossaries, document handling, and real volume enter the picture.
This comparison walks through how the major AI translation tools actually charge, why the billing units are not directly comparable, and where the hidden costs hide. One honest caveat up front: list prices in this market change often, and free-tier allowances have shifted repeatedly over the years. The figures below describe the pricing structures that have been documented, but anyone making a purchase decision should confirm current rates on each vendor’s official pricing page on the day they buy. The structures are stable; the numbers are not.
Why Headline Prices Mislead
The single biggest mistake buyers make is comparing a per-character rate against a per-token rate and assuming the smaller number wins. It does not work that way.
Vendors bill in at least three different units:
| Billing unit | Typical vendors | What it means in practice |
|---|---|---|
| Per character | DeepL, Google Cloud Translation, Azure AI Translator, Amazon Translate | You pay for the source text volume, usually per million characters |
| Per token | OpenAI and other large language model APIs | You pay for input and output tokens, which vary by language and model |
| Per word | Some freelance marketplaces and older tools | A rough proxy that rarely matches character counts |
A character-based rate looks transparent until you realize that a single Chinese character can carry the meaning of several English words, while a token-based model may split one English word into multiple tokens. Translation cost per source word therefore swings wildly depending on the language pair. Any comparison that skips this conversion step is guesswork dressed up as analysis.
The practical fix: build a small test corpus from your own content, run it through each candidate tool, and measure the actual billed amount. Ten pages of your real documents will tell you more than any pricing table.
The Major Players and How They Charge

DeepL
DeepL has offered a free consumer tier with character limits alongside paid Pro and API tiers since at least 2017 and 2018. Its API pricing has historically been structured per character, quoted as a rate per million characters. The free tier is genuinely useful for individuals testing quality, but it is a consumer product, not a production system.
A common mistake: teams sign up for the free tier, build a workflow around it, and only discover the character cap when a busy week hits the limit mid-project. Check the cap against your monthly volume before you commit to any architecture.
Google Cloud Translation
Google Cloud Translation has long offered two API tiers: the older Basic (v2) and the more capable Advanced (v3). Both are priced per million characters, and a free monthly character allowance has historically applied to the Basic tier, documented at 500,000 characters per month. The Advanced tier adds features like glossary support and batch translation, which matter for anyone translating branded terminology consistently.
The v2 versus v3 distinction confuses a lot of first-time buyers. If your content includes product names, legal terms, or slogans that must never be translated loosely, the Advanced tier’s glossary support is usually worth the higher rate.
Microsoft Azure AI Translator
Azure AI Translator has historically priced per million characters with a free tier offering 2 million characters per month on the F0 tier. Because it sits inside the broader Azure ecosystem, the real value for many organizations is integration: existing Azure billing, identity management, and security tooling apply without a separate vendor relationship.
The catch is that Azure pricing is a moving target. Free-tier allowances and per-character rates have been adjusted multiple times, so the number you remember from last year’s evaluation may no longer apply.
Amazon Translate
Amazon Translate has historically charged per million characters with a free tier of 2 million characters per month for the first 12 months. That 12-month framing is the detail people miss. It is a trial allowance, not a permanent free tier, and budgets built around it need revisiting after year one.
OpenAI and Other Language Model APIs
OpenAI’s API is priced per token, split between input and output, not per character. Translation cost therefore depends on how the model tokenizes each language, and tokenization varies significantly across scripts. Pricing has changed repeatedly since 2023, so any figure quoted in a comparison article should be treated as historical.
The upside of language model APIs is flexibility: you can control tone, formality, and terminology through prompts rather than paying separately for glossary features. The downside is that cost prediction is harder, because output length and token counts are less predictable than a straight character count.
Free Tiers Versus Paid Tiers: A Distinction Buyers Keep Missing
Articles about translation pricing routinely conflate two completely different products. The free web interface you use to translate a menu is not the same product as the paid API that powers a website localization pipeline.
Free consumer tools from DeepL, Google Translate, and Microsoft Translator are designed for human-in-the-loop use. They come with usage caps, no service guarantees, and no glossary or batch features on the free plan. Paid API tiers are built for automation: they offer higher limits, terminology control, and integration support.
The mistake to avoid is assuming a free consumer tier’s quality reflects the paid API’s capabilities, or vice versa. They often run on related but not identical systems, and the feature sets diverge sharply.
The Hidden Costs Nobody Puts on the Pricing Page
List prices capture only part of the real cost. Here is what tends to show up later.
Glossary and terminology management. Some vendors include glossary support in certain plans and charge for it in others. For brands with strict terminology rules, this can be the deciding factor.
Document handling. Translating a formatted document, a spreadsheet, or a slide deck is a different job from translating plain text. Format preservation features are sometimes bundled and sometimes billed separately.
Post-editing labor. Machine translation output still needs review for high-stakes content. The cheapest per-character rate can become the most expensive option if it produces output that takes twice as long to fix. Budget for human review time as a real line item.
Integration and engineering. API calls need code, error handling, and monitoring. If your team has no engineering capacity, a tool with a polished connector for your content management system may beat a cheaper raw API.
Volume commitments and enterprise discounts. Published list prices are starting points. Enterprise and volume discounts are typically negotiated privately, which means any comparison built purely on list prices understates how much pricing varies between buyers.
A Practical Comparison Framework
Instead of chasing the lowest headline number, run a structured evaluation. Here is a sequence that works.
- Define your monthly volume in characters. Pull real numbers from your content system rather than estimating.
- Identify your language pairs. Cost per word varies enormously between, say, English to Spanish and English to Japanese.
- List your must-have features. Glossaries, batch processing, document format preservation, and formality control are the usual four.
- Run a pilot on your own content. Translate 10 to 20 real documents with each finalist and review the output with the people who will actually use it.
- Calculate total cost, not unit cost. Add estimated post-editing hours and integration effort to the API bill.
- Confirm current rates in writing. Ask each vendor for a quote at your volume and check the official pricing documentation on the day you decide.
That last step matters more than it sounds. Free-tier allowances at Google, Azure, and AWS have all changed over time, and per-character rates have moved too. A comparison article, including this one, is a map of the terrain, not a live feed of prices.
Where AI Translation Fits in a Wider Compliance Picture
Pricing is not the only variable. Teams translating customer-facing content in the European Union should be aware that the EU AI Act, Regulation (EU) 2024/1689, came into force on 1 August 2024 and includes transparency obligations relevant to AI-generated content. This does not change what vendors charge, but it can affect which tools and workflows are acceptable for regulated content. Buyers in healthcare, finance, and public sectors should factor compliance review into the total cost of ownership, because legal review time is not free.
Common Mistakes That Inflate the Bill
A few patterns show up again and again in real procurement.
Buying on unit price alone, without modeling post-editing effort, is the most expensive mistake. A tool that costs half as much per character but requires heavy correction can easily double the effective cost.
Ignoring the difference between source and target character counts is another. Some vendors bill on source characters, others on output, and the gap can be substantial for languages that expand in translation.
Forgetting that free tiers expire is a third. The 12-month allowance on Amazon Translate and the monthly caps on consumer tools both create a cliff that catches teams off guard.
Finally, teams often skip the pilot. A one-hour test on five real documents would have prevented most of the disappointing purchases in this category.
Conclusion
The honest answer to which AI translation tool is cheapest is that it depends on your language pairs, your volume, your feature needs, and your tolerance for post-editing. Character-based APIs from DeepL, Google Cloud, Azure, and Amazon offer predictable costs and generous entry allowances, while token-based language model APIs offer flexibility at the price of forecasting difficulty. The vendors that look cheapest on a pricing page are not always the ones that cost least in practice once glossaries, document handling, and human review enter the equation.
The durable approach is to treat pricing pages as a starting point, run a pilot on your own content, and calculate total cost rather than unit cost. Confirm every rate directly with the vendor before signing, because this market moves fast and published numbers age quickly. Do that, and the invoice at the end of the quarter will hold far fewer surprises.
Frequently Asked Questions
Is per-token pricing cheaper than per-character pricing for translation?
There is no universal answer. Token-based billing depends on how the model splits each language into tokens, and that ratio varies by script and model. Character-based billing depends on source text volume. The only reliable comparison is to run the same sample content through both and compare the actual billed amounts.
What is the difference between a free translation tier and a paid API tier?
A free consumer tier is built for occasional human use, with usage caps and limited features. A paid API tier is built for automation, with higher limits, glossary support, batch processing, and integration options. They are different products, and quality or feature assumptions from one do not automatically carry over to the other.
Do free monthly character allowances still exist?
Free allowances have historically existed across the major cloud providers, but the amounts have changed multiple times. Some are permanent monthly allowances, while others, like a 12-month trial credit, expire. Always confirm the current allowance on the vendor’s official documentation before building a budget around it.
Why do two vendors quote such different prices for the same language pair?
Billing units differ, feature bundles differ, and some vendors include glossary or document handling in the base rate while others charge extra. Volume discounts also vary by buyer. Two quotes that look far apart on the surface often converge once you model the same feature set and volume.
Should I budget for human post-editing on top of the API cost?
Yes, for any content where accuracy matters. Machine translation output still needs review for legal, medical, marketing, and brand-sensitive material. Post-editing labor is often the largest line item in a translation budget, and ignoring it makes cheap tools look cheaper than they are.
How often do AI translation prices change?
Frequently. Per-character rates, free-tier allowances, and plan structures have all been adjusted multiple times in recent years, and token-based model pricing has changed repeatedly since 2023. Any published comparison should be treated as a snapshot, and buyers should verify rates directly with each vendor at the time of purchase.