DeepL

DeepL's free web translator did what its own pitch promises: every idiom we planted came out as an idiom, and every name, number, date and account code came out intact. Our limited 4/5 covers five automated anonymous fixtures on the web translator, not DeepL Pro, not document translation, and not the API.

What we tested. T1 sent an informal English paragraph to German with two idioms (touch base, green light), a name, a count and a weekday. T2 sent a formal German sentence to English with the idiom nicht mein Bier, a date and an amount. T3 sent an English request to French and then tried the Formality control. T4 sent an English sentence to Spanish carrying a 30-day window, a date, a time, a currency amount, a percentage and an account number. T5 pasted 1,765 characters to Spanish to see how the anonymous limit is handled. Each fixture was entered once, with no account and no sign-in.

The idioms. In T1, touch base became noch einmal darüber sprechen and green light became grünes Licht, an established German idiom; Priya, 14 and Donnerstag all survived, and the pronoun for the report (ihn) agreed with der Bericht. The one register choice we noticed is that DeepL introduced the informal dir, which matched the casual source. In T2, nicht mein Bier became none of my business, not a word about beer, and the salutation stayed formal: Dear Mr. Weber, the invoice dated March 3, 2026, for 1,250.00 euros is still outstanding, payment due by September 30.

The figures. T4 returned El plazo de 30 días finaliza el 12 de octubre de 2026 a las 17:30 h; la tasa asciende a 2.450,75 dólares estadounidenses (2,5 %), pagadera en la cuenta 0043-118. Every digit is present; the separators and the 24-hour clock were localised rather than copied. Tasa for fee is defensible; tarifa or comisión would also have worked.

The nuance. T3 returned Pourriez-vous m'envoyer le contrat signé d'ici vendredi ? Si vous avez des questions, n'hésitez pas à m'appeler. The vous register is right for a business request, but the second sentence became if you have questions, do not hesitate to call me, a natural French turn of phrase that is slightly warmer than the source's if anything is unclear, just call me. Nothing factual moved; the register did.

The gate. The Formality control in the side panel carries a Pro badge and is greyed out. Clicking it did nothing anonymously: no dialog, no upsell text, no options. The badge is the whole disclosure.

The limit. With 1,765 characters in the box the counter turned red at 1765 / 1500 and a notice appeared under the output: Want to translate longer texts? Only 1500 out of 1765 characters were translated. Sign up for free to translate up to 5000 characters at a time. That is clear disclosure. The output itself, however, simply stops where the source was cut, mid-word, ending Nuestro pequeño estudio st, with no marker at the cut. Copy the result without reading the notice and you ship a fragment.

Where this leaves the reader. For a paragraph of correspondence into or out of German, French or Spanish, the anonymous tool held every fact we checked and handled idiom the way a fluent writer would. The mid-word cut at 1,500 characters and the Pro-only formality toggle are the two places where the free web tool needs you to pay attention. The clean fidelity on four fixtures is why this scope earns 4/5 rather than 3/5; the unmarked truncation, one register shift and the thin five-fixture base are why it does not earn 5/5. That is an editorial judgment, not a pass count.

What this review cannot establish. Testing was automated on September 16, 2026 in an automated Chromium browser, anonymously, on four language pairs and five fixtures. DeepL did not disclose which model produced the translations. Our first attempt to insert T1 failed inside our own automation before any text reached the product, and is disclosed here rather than hidden. We did not test document translation, glossaries, DeepL Write, the API, the browser extension or apps, Pro plans, the signed-in free account's 5,000-character limit, or any other language. Correctness was checked against the fixtures' named tokens and our own reading of the output, not against a reference corpus. These gaps bound the review rather than lower the score, and no product-wide rating follows from them.