Best AI for translation in November 2026: 11 tools scored on 7 rules, with every source and date

Disclosure: NOXI publishes this page and is one of the eleven tools scored. That is why the rules, their points, the source of every cell and the date each page was opened are printed here, along with the two rules NOXI scores zero on and the five tools that finish ahead of it.

The interesting thing is not who won, but that the vendor with the most complete documentation came fourth.

Microsoft is the only one of these eleven whose own pages cover all four levels of the job: text, a whole file with the formats named, a stated promise that the layout survives, and right-to-left behaviour. It still lands on 86, because the rules also reward a published language count and a free tier whose page says no card is needed. Microsoft publishes neither. Smartcat publishes both, and that is its whole seven-point lead.

Where NOXI sits, plainly: sixth of eleven, on 53 points. Smartcat, DeepL, Google, Microsoft and Canva Translate finish ahead. NOXI translates a deck, not a document, which is exactly why rule T2 reads as it does. We scored ourselves on rules written for the document job without softening a cell.

Best AI for translation: the eleven tools at a glance

On seven documented rules read on 2 November 2026: Smartcat 93, DeepL 86, Google 86, Microsoft 86, Canva 61, NOXI 53.

The three-way tie at 86, and the rule that decided it. T1 breaks the first step: DeepL prints "100+" and Google "over 200 languages", worth 15 each, while no Microsoft page we could open states a total, worth 8. Microsoft is therefore the last of the three and is printed fourth. DeepL and Google have identical cell vectors, so table order decides it and a reader should treat them as level.

Who should pick what

Six jobs, five tools, one documented fact each.

• If you want to translate a real file for nothing today → Smartcat, whose document translator page offers "15,000 Smartwords ... no credit card." and names .docx, .pdf, .pptx and .xlsx [2].
• If the format list matters more than the price → DeepL, which names PDF, Word, PowerPoint, Excel, HTML, SRT and TXT and says it "preserves the formatting and layout of original designs" [3].
• If you only need a chunk of text in as many languages as possible → Google Translate, whose help page covers text, handwriting, photos and speech "in over 200 languages" [4].
• If the file has to come back looking like the original → Microsoft's document translation, the only row stating "Translate files while preserving the original layout and format." [5].
• If the thing to translate is a Canva design → Canva Translate, which documents "Mirror page to match text direction" and is free one time [6].
• If the thing to translate is a deck that must read right to left → NOXI, whose Translate a deck makes a separate Arabic version with mirrored timelines and tables and an editable PowerPoint download [1].

Key numbers

Four different jobs hide behind one question

Of the eleven scored, one covers all four levels on its own pages.

• Level one, text. Paste a paragraph, get a paragraph. Every tool here does this, so it separates nobody.
• Level two, a whole file, with the formats named. Four vendors name three or more formats including a presentation format. Four document a file-to-file job without naming three. Three document none.
• Level three, the layout. A translated file that arrives as a wall of text is a new job, not a finished one. Four state preservation outright, two make a weaker claim, and five say nothing.
• Level four, right to left. Arabic, Hebrew, Farsi and Urdu need more than a language code. They need the page to turn around, and four of the eleven document it.

That is why T2, T3 and T4 are separate rules, and why a tool can name forty formats and still score zero on the thing an Arabic reader notices first. For the deck job, our walkthrough of how to translate a PowerPoint covers the mechanics.

How we scored: rubric T, printed in full

Rubric T is new here and all 100 points are plain points. Three of the seven rules, worth 50 together, are about what happens to a file rather than to a sentence.

Four reading rules, so every row is scored the same way. T2 measures whole-document translation, not file creation, which costs Claude twenty points. T4 treats a writing-direction column in a language table as a fact about the language, not a product claim, which is why Smartcat scores 8. T5 needs all three of its parts on one page, which takes seven points off Microsoft and gives fifteen to NOXI. T7 counts only surfaces we could read.

Every cell, the arithmetic and the ties

Eleven tools, seventy-seven cells, every total the plain sum of its row.

The two ties, and the rule that broke each one

The tie at 86 is broken by T1, which sends Microsoft to fourth. DeepL and Google match on all seven cells, so table order prints DeepL above Google. Lokalise and Phrase tie on every cell too, so table order decides again. Two placings rest on nothing but row order, and both are labelled where they sit.

Gap lines, between the neighbours worth comparing

• Smartcat 93 minus DeepL 86 = 7, and the whole gap is T5. Smartcat says "no credit card"; DeepL publishes 50,000 characters a month and says nothing about one.
• Google 86 minus Microsoft 86 = 0, made of T1 plus 7 and T4 minus 7, which cancel. Google prints a language total and no right to left; Microsoft the reverse.
• Microsoft 86 minus Canva 61 = 25, made of T2 plus 10, T3 plus 7, T6 plus 10 and T7 plus 5, minus T1 at 7.
• Canva 61 minus NOXI 53 = 8, made of T1 plus 15, minus T5 at 7. Canva prints a language count and NOXI prints none.
• NOXI 53 minus Gamma 35 = 18, made of T2 plus 10 and T3 plus 8.

Sensitivity: three readings, printed whether they help us or not

• Equal weights, each cell as a fraction of its own maximum: Smartcat 0.933, DeepL 0.867, Google 0.867, Microsoft 0.867, Canva 0.581, NOXI 0.505, ChatGPT 0.371, Lokalise 0.362, Phrase 0.362, Gamma 0.357, Claude 0.219. Only Gamma moves, from seventh to ninth.
• Without T2, the heaviest rule, out of 80: Smartcat 73, DeepL 66, Google 66, Microsoft 66, Canva 51, NOXI 43, Gamma 35, ChatGPT 34, Lokalise 23, Phrase 23, Claude 18. The order does not change.
• With T4 dropped, out of 85: Smartcat 85, DeepL 78, Google 78, Microsoft 71, Canva 46, NOXI 38, Lokalise 33, Phrase 33, ChatGPT 26, Gamma 20, Claude 18. That is the honest reading for a team where nobody reads right to left: Microsoft falls further and Gamma drops to tenth.

Best AI for translating documents: which tools name the formats

Four of the eleven name three or more document formats including a presentation format, the only route to full marks on T2.

Smartcat is the most specific: "Supports .docx, .pdf, .pptx, .xlsx and more", plus "Tables, columns, fonts, images, headers, footers and numbering stay where they were." [2] DeepL names the most formats in one page of prose, PDF, Word, PowerPoint, Excel, HTML, SRT and TXT, with its free version capped at one file a month and 5 MB [3]. Google moved from 61 to 86 on a page nobody had opened for the first draft: Cloud Translation lists DOC, DOCX, PDF, PPT, PPTX, XLS and XLSX and "preserves the original formatting and layout in your translated documents" [4]. Microsoft puts it in one line, "Translate files while preserving the original layout and format." [5]

Canva Translate, Lokalise and Phrase document a file-to-file job without clearing the three-format bar, so each scores 10. NOXI takes the same 10 for a different reason: Translate a deck takes a deck in and returns a translated deck as an editable PowerPoint file or a PDF, which is two formats out and a single object type in. Gamma, ChatGPT and Claude document none at all, which is the largest correction in this table: Claude creates .xlsx, .pptx, .docx and PDF files, a different thing from translating one, so T2 is zero and the total 18 rather than 38.

Does the layout survive? What each vendor actually claims

Four of the eleven state preservation outright, two make a weaker claim, and five say nothing on any page we could open.

This is the rule most roundups skip, and it decides whether you spend Sunday repairing a file.

Full marks, 15 points. Microsoft: "Translate files while preserving the original layout and format." Google: "preserves the original formatting and layout in your translated documents". DeepL: "preserves the formatting and layout of original designs". Smartcat is the most detailed: "Tables, columns, fonts, images, headers, footers and numbering stay where they were."

Half marks, 8 points. Canva's help page makes no preservation claim. It documents "Mirror page to match text direction", and a mirroring claim is not a preservation claim [6]. NOXI is the other 8, with the caveats printed rather than tidied away: mirroring is documented, charts are deliberately not mirrored with the single exception of funnels, and Arabic text in a PowerPoint download can differ slightly from the browser [1].

Zero. Gamma, ChatGPT, Lokalise, Phrase and Claude make no layout claim we could find. Gamma goes the other way on a related question: "Charts imported from Google Sheets export to PowerPoint and Google Slides as static images" [7]. One phrase appears nowhere on this page about any tool, ours included: nobody in this field says a chart stays editable after it leaves the tool.

Right to left, and why it is a separate question from Arabic

Four of the eleven document right-to-left behaviour: Microsoft, Canva, Gamma and NOXI. Four more name Arabic and stop there.

This distinction is the most useful thing here. "Arabic supported" means the engine has a language code. Right to left means the product turns the page around: the columns swap and the timeline runs from the right.

Full marks, 15 points. Microsoft's Office article says "Microsoft 365 supports right-to-left functionality and features", and PowerPoint offers the only dialect choice anyone here documents, "change it to an Arabic dialect" [5]. Canva documents "Mirror page to match text direction" [6]. Gamma says "Gamma supports right-to-left (RTL) languages, including Arabic, Farsi, Hebrew, and Urdu" and answers the formatting question honestly in the same article, "Yes, but formatting may require manual adjustments." [7] NOXI has a full Arabic interface, right-to-left slides, mirrored timelines, Gantt charts, chevron processes and direction-aware tables, and four Arabic fonts [1].

Eight points, Arabic named only. Smartcat's languages table lists Arabic with the tag ar, the writing direction "right-to-left" and thirteen regional variants, a fact about the language rather than a claim about a layout. DeepL's help centre names "Modern Standard Arabic (MSA) available as a source and target language". Google lists ar and ar-SA as officially supported, and ChatGPT's interface language article includes Arabic with no total.

Zero. Lokalise's languages page returned 404 at 14:36 and a Phrase supported-languages article returned 404 at the same minute, so neither names Arabic on a page we could open. That is the basis of the cell, not a claim that either lacks Arabic. Claude's pages document multi-language support without naming Arabic or right to left.

If Arabic is the requirement rather than a nice-to-have, our ranking of the best AI for Arabic scores tools on rules written for it, and building a bilingual Arabic and English deck covers the design side.

Free tiers, with the numbers the vendors publish

Three take full marks on T5, and what separates them is one sentence about a card.

Prices, in the currency each page served us. DeepL served euros, with a selector reading Germany, so the only paid figure we can print is the one it showed, "€7.49" per month for Individual billed annually. Smartcat prints "from $1,200 / yr" for Adapt and names no currency code anywhere on its pricing page, so the figure is printed as the page writes it and nothing is converted. From Microsoft's Azure Translator pricing page we print one figure, "$40 per million characters of document translation" on instance D3, and that page does state that prices are calculated in US dollars. Phrase prints no currency symbol beside its plan numbers, so no Phrase price appears here, and Google's Cloud Translation pricing page did not render its content to us, so no Google price appears either.

Glossaries and house terminology

Six of the eleven document a glossary or terminology control, and they are not the six you would guess.

If your product names or legal entities have to come out the same way every time, this rule decides your tool. The six are Smartcat, DeepL, Google, Microsoft, Lokalise and Phrase. Two of them finish ninth and tenth overall and still hold this cell, which is the clearest sign that a total is not a recommendation.

Page by page: Smartcat says "Every job uses your glossary, style guide and translation memory". DeepL's developer documentation reads "A dictionary is a mapping of source phrases to target phrases for a single language pair." Google says "A glossary is a custom dictionary the Cloud Translation API uses". Microsoft says "Dictionaries are sometimes called glossaries or term bases." [2] [3] [4] [5]

Five score zero: Canva Translate, Gamma, ChatGPT, Claude and NOXI. There is no glossary and no terminology control anywhere in NOXI's documentation, so the cell is zero and stays zero. If house terms matter in a deck, put them in the brief: a brief is read in full and a translated deck keeps the names it is given.

Subtitles, in one section

One of the eleven names a subtitle format on its own pages: DeepL, with SRT in its document list.

An SRT is a plain-text file of numbered caption blocks with timecodes, so any tool that translates TXT can in principle translate one. Naming it is not the same as doing it well, and this rubric scores the naming.

What the seven rules miss: reading speed and characters per line, whether timings shift when the target language runs longer, speaker labels, burned-in captions against a sidecar file, and the formats broadcasters ask for.

NOXI: sixth of eleven, and what the 53 is made of

NOXI scores 53 of 100 and finishes sixth, with full marks on two rules, two zeros and three cells in between: the shape you would expect from a deck tool measured against document translators. Translate a deck is one of the workflow actions on Home for every account, next to Create from file, Beautify my deck, Improve structure and Match reference.

Two facts the rules do not score. The translated deck is new rather than a linked pair, so the original stays beside it. And the export writes real text boxes with the font programs embedded, Noto Sans Arabic included, which stops an Arabic deck arriving in the wrong typeface on someone else's laptop: see our guide to embedding fonts in PowerPoint.

NOXI is also listed in ChatGPT's plugin directory, so you can ask for a deck in a chat and open it in NOXI to edit or export. Availability depends on your plan, workspace, role and region. See our write-up of the NOXI plugin for ChatGPT. For the region, our guide to the best AI presentation maker in the UAE covers it.

One tender, two languages: Yousef Al Marri's Sunday deadline

Yousef Al Marri, a bid manager in Doha, has one deadline and two files, and the answer is two tools.

He is fictional, the arithmetic is not. On Wednesday the client asks for the submission in Arabic by Sunday: a 60-page English tender response and a 24-slide summary deck.

Counting the document. Sixty pages at about 550 words a page is 60 × 550 = 33,000 words. At six characters a word including spaces that is 33,000 × 6 = 198,000 characters. Now the free tiers mean something. Microsoft's Azure allowance of 2 million characters a month covers it with room to spare: 198,000 ÷ 2,000,000 = 0.099, so 9.9 per cent of one month. DeepL's free version would need 198,000 ÷ 50,000 = 3.96, four months of allowance, and caps him at one file a month anyway. Against Smartcat's 15,000 free Smartwords his 33,000 words are 33,000 ÷ 15,000 = 2.2 times the allowance. At the document translation rate the Azure pricing page prints for instance D3, $40 per million characters of document translation, the file would cost 198,000 ÷ 1,000,000 × 40 = $7.92.

Choosing for the document. He needs the layout back, and Microsoft is the only row that states preservation and also documents right to left and an Arabic dialect choice in Office. At 15,000 words Smartcat would be the cheaper route at zero; at 33,000, Microsoft's free allowance decides it.

Choosing for the deck. He opens it in NOXI, chooses Translate a deck, picks Arabic, and gets a separate Arabic version: slides running right to left, timelines and tables mirrored, four Arabic fonts in the design panel. He downloads it as an editable PowerPoint file with the fonts embedded, for 25 of his 1,000 monthly credits: 25 ÷ 1,000 = 2.5 per cent of the month.

The 42 product names, and the bill. His catalogue has 42 names that must be identical in both files. Microsoft documents dictionaries, so the tender route holds them; NOXI documents no glossary, so he pastes the 42 terms into the deck brief and checks them at the end. Two files, two free allowances, $0, reviewed on 1 December 2026.

What these rules don't measure

Seven rules cannot hold a category, and some of what they leave out is why people pay for these tools.

• Translation quality, judged by a native reader. Nothing here scores whether the Arabic is good. Smartcat, DeepL, Google and Microsoft all have decades of linguistic work this table cannot see.
• Dialects in practice. Microsoft takes the dialect point because it documents the choice in Office. How any engine handles Gulf, Levantine or Egyptian Arabic is a question for a linguist.
• Human review, certification and turnaround. Smartcat, Lokalise and Phrase are built around a workflow with people in it. The last two finish ninth and tenth here and are what a localisation team would buy, because these rules ask about files and those products are pipelines.
• Speech, voice and handwriting. Google documents all three on the same page as its 200 languages, and none of it is scored.
• Legal and regulated accuracy, translation memory and cost per word at volume.
• Data residency, retention and who sees the file. A tender under embargo is a procurement question first.

What changed recently on the vendors' pages

Five dated movements, read on 2 November 2026. Two of them changed a score here.

Two pages did not render, and nothing is printed from either. Google's Cloud Translation pricing page returned only its title and a truncation notice at 14:37, on two attempts. Azure's Translator marketing page timed out at 60 seconds at 14:30, so Microsoft's language cell rests on documentation pages that state no total.

What would change this ranking

Three things would move a row, each tied to a page and a re-check date.

1. DeepL documents right to left anywhere on its own pages

DeepL names the widest format list here and scores 8 on T4 because its help centre names Modern Standard Arabic and documents no right-to-left behaviour. One right-to-left article would take T4 to 15 and DeepL to 93, level with Smartcat, where the chain ties on T1 and T3 and then reaches T5, where Smartcat's 15 keeps first place and DeepL takes second outright, ahead of Google and Microsoft. The pages to watch are support.deepl.com, which opened to a logged-out browser at 14:34 on 2 November 2026, and deepl.com's feature pages. We re-check on 1 December 2026.

2. Microsoft's Azure pricing page states that the free tier needs no card

Microsoft loses seven points on one absence. Its pricing page publishes a free monthly allowance of 2 million characters and says nothing about a credit card, which is the printed 8 level. Add that sentence and T5 goes to 15, Microsoft reaches 93, and the tie with Smartcat is decided on T1, where Smartcat's 15 against Microsoft's 8 keeps Smartcat first and lifts Microsoft to second. The page is azure.microsoft.com's Translator pricing details, read at 14:25 on 2 November 2026. We re-check on 1 December 2026.

3. NOXI publishes a supported-language count

Our own zero on T1 is the easiest cell here to change, and it is ours to change. No NOXI page states how many languages a deck can be written or translated into, so the cell is 0 rather than a figure we have not published. Put a count of 100 or more on noxi.ai and T1 goes to 15, NOXI reaches 68 and moves from sixth to fifth, ahead of Canva Translate's 61. We re-check noxi.ai on 1 December 2026.

Build it: translate a deck so it reads right to left

This is the deck half of Yousef's job, using the Translate a deck action on Home. The translation is a new deck rather than a linked copy, so the original stays as it was, and the numbers and names travel from the source deck rather than being fetched.

Two sister pages pick up where this one stops: our ranking of the best AI for reports scores written reporting on the same kind of rules, and our guide to a disability awareness presentation shows the brief-and-review pattern where wording carries weight.

Frequently asked questions

What is the best AI for translation?

On seven printed rules checked on 2 November 2026, Smartcat scores highest with 93 of 100. DeepL, Google and Microsoft all land on 86, Canva Translate takes 61 and NOXI is sixth of eleven on 53. Microsoft is the only one of the eleven documenting all four levels of the job.

What is the best AI translator?

For a chunk of text, Google Translate documents "over 200 languages" and costs nothing. For a whole file, Smartcat's document translator page publishes 15,000 free Smartwords and says no credit card. Both were read on 2 November 2026. Smartcat leads our table on 93 points, Google ties DeepL and Microsoft on 86.

What is the best AI translation tool?

It depends on the format. Smartcat scores 93 of 100 because it names .docx, .pdf, .pptx and .xlsx, states that tables, fonts, headers and footers stay in place, and publishes a free allowance with no card. DeepL, Google and Microsoft follow on 86, checked 2 November 2026.

What is the best AI for document translation?

Four of the eleven name three or more document formats including a presentation format: Smartcat, DeepL, Google and Microsoft, all read on 2 November 2026. Microsoft's Azure page is the one that says files are translated "while preserving the original layout and format". Smartcat is first overall on 93.

What is the best AI for translating PDF files?

DeepL names PDF first in its document list and Google's Cloud Translation page accepts PDF alongside DOC, DOCX, PPT, PPTX, XLS and XLSX, both read on 2 November 2026. Canva has a PDF translator page of its own, scoring 61, with a 2000 character limit per use.

What is the best free AI translator?

Smartcat publishes 15,000 Smartwords with no credit card, DeepL 50,000 characters a month and one file, Microsoft 2 million characters a month on Azure, and Gamma and NOXI both publish a free allowance with no card. All five pages were read on 2 November 2026.

What is the best AI for subtitles?

DeepL is the only one of the eleven whose document-translation page names SRT, the subtitle format, read 2 November 2026. Nobody else in the table names it. Timing, speaker labels and burned-in captions are outside these seven rules, so treat the table as a document answer rather than a video answer.

DeepL or Google Translate, which is better?

They are level. On 2 November 2026 both score 86 of 100 with identical cells: 15, 20, 15, 8, 8, 10, 10. Nothing in the rubric separates them, so the printed order falls to table order. DeepL served euros to us; Google's pricing page did not render.

Which AI translates a deck into Arabic and keeps it right to left?

Four of the eleven document right to left on 2 November 2026: Microsoft, Canva, Gamma and NOXI. For a deck, NOXI's Translate a deck makes a separate Arabic version with mirrored timelines and tables and four Arabic fonts, and the download is an editable PowerPoint file with the fonts embedded.

Does any AI keep a document's formatting after translation?

Four vendors state it outright on 2 November 2026. Microsoft writes "while preserving the original layout and format", Google "preserves the original formatting and layout", and DeepL "preserves the formatting and layout of original designs". Smartcat adds that tables, fonts, headers and footers stay where they were.

Sources & methodology

Every vendor page below was opened on 2 November 2026, logged out, with no cookie banner accepted and no sign-in. Three paths were used: a command line fetch for status codes, a server-side fetch for page text, and a logged-out browser for pages that refuse the other two. Times are local, UTC plus four, in a window of 14:16 to 14:37. Prices are printed only in the currency each page served, which is why there is no dollar figure for DeepL, no currency beside a Phrase plan and no Google price. Scores come from our shared November 2026 score file and no weight was moved to place any row. Yousef Al Marri is fictional; his arithmetic is computed from the figures above.

• [1] NOXI capabilities, Arabic and right-to-left behaviour, export, credits and the connector · noxi.ai and NOXI product documentation, rows 2, 9, 10a, 12, 13 and 16e; read 2 November 2026, 14:16.
• [2] Smartcat · Document translator (14:20); Pricing (14:25, plan figures printed with a bare dollar sign and no currency code); Languages (14:36); checked 2 November 2026.
• [3] DeepL · Document translation (14:25); Pro and free version (14:25, served euros); Help centre (logged-out browser, 14:34 to 14:35); Glossaries API reference; checked 2 November 2026.
• [4] Google · Cloud Translation document translation (14:31); Google Translate help; Supported languages; Glossaries; the pricing page returned only its title and a truncation notice at 14:37, on two attempts, so no Google price is printed; checked 2 November 2026.
• [5] Microsoft · Azure document translation overview (ms.date 7 August 2026); Azure Translator overview (14:32, no language total on the page); Right-to-left languages in Office (14:31); Translator pricing details (14:25); Custom Translator dictionaries (14:20); checked 2 November 2026.
• [6] Canva · PDF translator (134 languages) (14:20); Translate Canva designs (100+ languages, mirror page) (14:25); checked 2 November 2026. The two language counts disagree and both are printed above.
• [7] Gamma · Does Gamma support right-to-left languages; Pricing (a bare dollar sign with no currency code); the export article on charts from Google Sheets, re-read at 14:34; checked 2 November 2026.
• [8] OpenAI · Language setting in ChatGPT (14:27); Creating and editing documents with ChatGPT Work (14:29, updated 1 November 2026); Uploading files and audio (14:28, updated 1 November 2026); checked 2 November 2026.
• [9] Lokalise and Phrase · Lokalise pricing; Phrase pricing (no currency symbol beside the plan numbers); the Lokalise languages page and a Phrase supported-languages help article both returned 404 at 14:36; checked 2 November 2026.
• [10] Anthropic · Claude pricing (14:18; plan figures carry a bare dollar sign and only the Enterprise line names US dollars, and no Claude price is printed here); Create and edit files with Claude (14:18); no Anthropic page we opened documents translation, a language count, a glossary or right to left; checked 2 November 2026.

Send the document to the tool that names your format and promises your layout back, and translate the deck in NOXI: free, right-to-left slides, four Arabic fonts, and an editable PowerPoint file with the fonts embedded.