The AI Cost Lab

What it actually costs to translate a landing page with AI

Three tools, three billing models, and a 200× price gap that no comparison table shows.

Here is a question that almost nobody answers with an actual number: what does it cost to translate one 1,500-character product page into five languages?

Ask three vendors and you get three answers that cannot be compared. One sells you a monthly subscription. One sells characters. One sells tokens, a unit that does not appear anywhere in the task you are trying to do. So people pick whichever pricing page they landed on first, and quietly overpay by an order of magnitude.

I ran the arithmetic. For a single page, the cheapest legitimate route costs about ten cents. The most common choice costs twenty dollars a month. Same output, same task.

The task, stated precisely

One Chinese product page, 1,500 characters. Translated into five languages: English, Spanish, German, Japanese, French. Source stays constant so the comparison is fair.

1,500 Chinese characters is roughly 900–1,000 English words per language, depending on the pair. That is a real page, not a toy example.

Three billing models that look similar and are not

DeepL — charged per character, capped per month

DeepL Pro prices on the volume of text you push through it. The free tier allows 50,000 characters a month and one document translation. Individual is $10.49/month, or $8.74/month billed annually, and raises the ceiling to 300,000 characters with three document translations.

Our job is 1,500 characters × 5 languages = 7,500 characters. That is 15% of the free monthly allowance.

Cost: $0.

ChatGPT and Claude subscriptions — a flat fee, decoupled from the task

ChatGPT Plus is $20/month. Claude Pro is $20/month, or $17/month billed annually. Neither has a character meter. You get a usage allowance measured in messages and rolling time windows, and the meter resets whether you use it or not.

This is the pricing model almost everyone defaults to, because it is the one you can buy with a credit card in ninety seconds and it comes with a chat box.

Cost for this job: $20.

API access — charged per token, no subscription

Claude Sonnet 5 through the API costs $2 per million input tokens and $10 per million output tokens. Chinese text runs roughly 1.3–1.7 tokens per character, so 1,500 characters is about 2,000–2,500 input tokens per language.

• Input: ~12,500 tokens × $2/MTok = $0.025
• Output: ~7,000 tokens × $10/MTok = $0.07
• Total: under $0.10

Side by side

Read that table again. The most expensive options are not the most capable. They are simply the ones that do not require you to know what a token is.

Where the subscription actually wins

I am not arguing that the $20 plans are a scam. They win in three specific situations, and it is worth knowing precisely which ones.

You use the tool for other things. That is the honest case for Plus or Pro. If you also write copy, debug, draft emails, and analyse spreadsheets with it, the translation job is a rounding error inside a tool you would pay for anyway. The comparison above is unfair to subscriptions, and it is the only fair way to show the unit cost.

You cannot hold a spreadsheet of tokens in your head. API billing is metered, and metered billing makes people anxious. The premium you pay for a flat fee is, functionally, an insurance premium against your own uncertainty. That is a legitimate thing to buy.

You are doing something the API cannot do. Document translation with preserved layout is the clearest example. DeepL's file translation keeps tables, headers, and formatting intact — PDF, Word, PowerPoint. Rebuilding that through an API is a project, not a task.

The cost nobody puts in the table

Every number above is wrong in the same way, because none of them includes the only cost that actually matters: the time it takes to make the output usable.

Machine translation of a marketing page is never publishable as-is. The German version will be grammatically perfect and tonally wrong. The Japanese will be over-direct in a way that reads as rude. Placeholders, product names, and idioms break in ways that no character counter sees.

A realistic estimate for reviewing five languages of a 1,000-word page: two to four hours, assuming you can actually read all five languages.

Human review time per page × 5 languages. At $50/hr that is $100–200 — against $0.10 of machine time.

This reframes the whole decision. Choosing between a $8.74 plan and a $20 plan is optimising the wrong digit. The real lever is reducing review time: tighter glossary control, better source copy, and fewer target languages that you cannot personally evaluate.

Which is also why "cheaper per character" is a bad buying criterion. The tool that produces output needing forty minutes of cleanup beats the one that produces output needing ninety, at almost any price.

What I would actually do

1. Under ten pages a month: DeepL Free, and spend the $20 on something else. If the free character limit bites, Individual at $8.74 is the cheapest sane upgrade.
2. Ten to fifty pages a month: the API. Ten cents a page does not become wrong until you are pushing volumes where the engineering time to build the pipeline exceeds the savings. It almost never does at this scale.
3. Anything with documents, layouts, or shared terminology: DeepL Pro Team, because formatting preservation is worth more than the difference in per-character rates.
4. Everything you already use daily: whatever you are already paying for, and stop pretending the translation was a separate purchase.

On the numbers

All prices were read from vendor pricing pages on November 2, 2026, in USD unless noted, before tax. Token estimates are estimates — tokenisers differ by model, and a 15% error either way changes the API line by a cent and a half, which is the point. Subscription prices change without notice and several vendors now run three or more tiers; check the source page before you commit to a year.

DeepL charges per character on the source text, and its own free-tier figure has been reported inconsistently across third-party sites — verify against the live pricing page rather than any comparison article, including this one.