productivity

Best AI for Translation

Compare the best AI translation tools and models as of July 2026 — Google Translate on Gemini 3.5, DeepL, ChatGPT, Claude, Qwen and Qwen-MT, plus dedicated MT APIs and localisation platforms, with quality benchmarks, language coverage, pricing and picks for every use case.

Updated July 2026

Quick answer: For everyday translation in mid-2026, the best free tool is Google Translate, now rebuilt on Gemini 3.5 and covering 249 languages (Google). For nuanced, professional text where tone and context matter, a frontier chat model — GPT-5.6 Sol via ChatGPT, Claude Opus 4.8 via Claude, or Gemini 3.5 Pro — now edges out dedicated engines on human evaluation for high-resource languages. DeepL remains the safest pick for formatted business documents, glossaries and formality control, and Qwen3.7-Max (and its dedicated Qwen-MT API) is the strongest for Chinese, Japanese, Korean and other Asian languages. For real-time spoken translation, Gemini 3.5 Live Translate handles speech-to-speech in 70+ languages.

The honest answer depends on what you’re translating, which languages you need, and whether you want a free app, a paid API, or a full localisation pipeline. This guide covers all four — general-purpose LLMs, dedicated machine-translation (MT) engines, consumer apps and localisation platforms — with current quality evidence, language coverage and pricing. The single biggest shift since our last update: general-purpose LLMs have overtaken purpose-built translation engines on human-rated quality for high-resource language pairs, while dedicated tools keep the edge on formatting, glossaries and cost at scale.

The current state of AI translation: July 2026

Machine translation crossed a threshold in the past year: the best general-purpose models are now, on human evaluation, as good as or better than the engines built specifically to translate.

The evidence comes from the field’s most rigorous public test. The WMT25 General Machine Translation shared task (Tenth Conference on Machine Translation, November 2025) evaluated 60 systems across 30 language pairs — 36 participant submissions plus 24 translations collected from large language models and popular online providers — using professional annotators and deliberately harder, document-level test sets. The organisers titled their findings paper “Time to Stop Evaluating on Easy Test Sets,” a signal that the easy wins are over and LLMs now compete head-to-head with dedicated engines (ACL Anthology). Independent 2026 round-ups reach the same conclusion: LLMs such as GPT-5.x, Claude and Gemini have overtaken specialised engines for high-resource pairs on human evaluation, though the gap narrows for low-resource languages and domain-specific content (intlpull).

The market reflects the shift. Industry forecasts put the AI language translator tool market at roughly $9.49 billion in 2026, rising to about $57 billion by 2035 (22% CAGR) (Precedence Research), while the narrower machine-translation-engine market sits near $1.26 billion (Mordor Intelligence). Cost has collapsed alongside: the total cost per word for AI-assisted localisation fell from roughly $0.20 to $0.002 across 2025–2026 — a 100x reduction — moving AI translation from a nice-to-have to the default for most teams (withallo). LLMs now reach an estimated 85–90% of professional human-translator quality for most content types at 5–10% of the cost (GetBlend).

Five shifts define the current moment:

1. LLMs overtook dedicated MT on human quality. For high-resource pairs (English–Spanish, English–German, English–Chinese), frontier chat models now match or beat DeepL and Google’s classic engine on human evaluation and context-sensitive metrics. The advantage is nuance: idioms, register, tone and document-level consistency.

2. The metric you cite decides the winner. Dedicated engines still lead on BLEU — the older surface-overlap metric — especially for short, formulaic strings like UI labels and error messages. LLMs lead on COMET (a learned, semantic metric) and on human judgement, which COMET tracks far better than BLEU (intlpull). A single “best” claim is meaningless without naming the metric.

3. Google Translate became an LLM product. Google rebuilt Translate’s text quality on Gemini and added Gemini 3.5 Live Translate for real-time speech, so the most-used free tool on earth is now a frontier-model product rather than a legacy engine (Google).

4. Dedicated engines specialised into workflow. DeepL, Qwen-MT and the cloud APIs increasingly compete on document formatting, grammatically-aware glossaries, formality control, data residency and predictable per-character pricing — the things a raw chat model does not give you out of the box.

5. Character pricing versus token pricing split the market. NMT APIs bill per character and stay predictable at volume; LLM APIs bill per token and cost 3–7x more per unit of translated text, but bring reasoning, context and customisation. Above roughly 10 million characters a month, character-priced engines are usually cheaper (SimpleLocalize).

Top AI for translation (July 2026)

No single tool wins every job. The table ranks the leaders by overall standing across quality, language coverage, price and workflow fit; the sections below explain each pick and who it suits.

Reading the price column: character prices (per million characters) and token prices (per million tokens) are not directly comparable — one million characters of English is very roughly 250,000 tokens, and LLMs also bill for output. Translating a million characters through an LLM API typically costs several times a character-priced NMT engine; see the pricing section for a like-for-like view.

Quality: which is actually most accurate?

There is no single accuracy champion — it depends on the metric, the language pair and the content type.

• For nuanced, context-dependent prose (marketing, literature, dialogue, register-sensitive text), frontier LLMs lead. GPT-class models have historically topped COMET and human evaluation, with Claude close behind and Gemini strong, particularly on document-level consistency (GetBlend).
• For short, formulaic strings (UI labels, error messages, product data), dedicated engines like DeepL and Google’s classic model still win on BLEU and are cheaper to run at scale (intlpull).
• For Chinese, Japanese and Korean, Qwen models handle honorifics, keigo and measure words more reliably than general Western-trained systems, and Naver’s Papago is the strongest consumer app for the same languages (RoundTalk).
• DeepL’s own blind tests rate its next-gen model 1.7x higher than its previous model for English→Japanese and English→Chinese and claim it beats GPT-4, Google and Microsoft for enterprise quality (DeepL). Treat this as a vendor claim: independent human-evaluation round-ups now place frontier LLMs at or above DeepL for high-resource pairs, while DeepL keeps its edge on formatting-preserving document workflows and formulaic content.

Best AI translation tools compared

Ordered by how most users will encounter them — free apps first, then the paid models and engines, then the localisation platforms that wire translation into software.

1. Google Translate — best free everyday translator

Price: Free (app and web); Google Cloud Translation API $20/M characters, first 500,000 free Languages: 249 languages and varieties Best for: Quick everyday translation, travel, 100+ language reach, spoken translation

Google Translate is the default for a reason: it is free, covers 249 languages and language varieties (as of April 2026), and Google has rebuilt its text quality on Gemini, so translations now handle idioms and slang far more naturally than the old phrase-based system (Google). The new Gemini 3.5 Live Translate model adds near real-time speech-to-speech translation in 70+ languages, preserving the speaker’s intonation, pacing and pitch (WinBuzzer). Camera, conversation and offline modes round out the widest free feature set available.

Limitations: The very best nuance still goes to a full chat model you can prompt for tone and context; the Gemini-quality text upgrade is rolling out by region and language pair rather than everywhere at once (Slator).

Best for: Anyone who wants a free, broad, capable translator for daily use and travel.

2. ChatGPT (GPT-5.6 Sol) — best for nuanced, in-workflow translation

Price: Free tier; ChatGPT Plus $20/month; API billed per token Models: GPT-5.6 Sol (current flagship, GA 9 July 2026); GPT-5.5 (prior flagship; GPT-5.5 Instant is the free/Go default) Best for: Context-aware translation you can direct — tone, audience, glossary, explanation

ChatGPT is the most flexible translator because you can instruct it: specify the register, the audience, the regional variant, or ask it to keep a brand glossary and explain its choices. GPT-class models have led COMET and human evaluation for nuanced, context-dependent translation (GetBlend), and unlike a fixed engine, ChatGPT can translate, localise, and adapt in one pass — rewriting idioms for the target culture rather than rendering them literally.

Limitations: No guaranteed formatting preservation for complex documents, per-token pricing is dearer than NMT engines at volume, and quality drops on genuinely low-resource languages where training data is thin.

Best for: Professionals translating marketing copy, correspondence and nuanced text who want to steer the output.

3. Claude (Opus 4.8 / Sonnet 4.6) — best for long documents and tone

Price: Free tier; Claude Pro $20/month; API Sonnet 4.6 $3/$15, Opus 4.8 $5/$25 per million tokens Best for: Long-document consistency, faithful tone, careful register

Claude is the strongest pick when a whole document must stay consistent — the same term translated the same way on page 1 and page 40, the author’s voice preserved throughout. Its 1M-token context window lets you translate book-length material in one pass without losing the thread, and reviewers consistently rate Claude highly for faithful, natural tone. Opus 4.8 is the quality pick; Sonnet 4.6 is the cost-effective one for high volume.

Limitations: Like all chat models, it won’t preserve intricate document formatting automatically, and it trails Qwen on some Asian-language specifics.

Best for: Translators of long-form content — reports, books, contracts — who prioritise consistency and voice.

4. Gemini 3.5 Pro — best for multimodal and live translation

Price: Via Google AI Studio / Gemini API (limited rollout); free tier in the Gemini app Best for: Image and document translation, and real-time spoken translation

Gemini 3.5 Pro is Google’s flagship and the engine now behind Google Translate’s best output. Its strengths for translation are multimodal input (translate text inside images, screenshots and PDFs), a very large context window for long documents, and the Gemini 3.5 Live Translate speech model for natural real-time conversation across 70+ languages (Google).

Limitations: Flagship availability is still a limited rollout; for pure text, a well-prompted GPT-5.6 Sol or Claude Opus 4.8 is an equally strong alternative.

Best for: Anyone translating images, mixed-media documents or live speech inside the Google ecosystem.

5. DeepL — best for formatted business documents

Price: Free tier (500,000 characters/month); paid plans from about $10/month; API $25/M characters Languages: 30+ core languages (135 source / 143 target counting variants) Best for: Enterprise document translation, glossaries, formality control, data governance

DeepL built its reputation on natural output and has kept a genuine workflow edge that raw chat models lack. Its document translation preserves Word, PowerPoint and PDF formatting; its glossaries adapt grammatically rather than doing crude find-and-replace; and it offers formality toggles and an LLM-powered “Clarify” feature for disambiguation (DeepL). DeepL’s next-generation model is itself LLM-based, trained on seven years of proprietary data, and Pro users can switch between the next-gen and “classic” models per job.

Limitations: Narrower language coverage than Google (33 core languages versus 249), and independent human-evaluation round-ups now rate frontier LLMs at or above DeepL for high-resource prose — its clearest remaining advantages are formatting, glossaries and predictable pricing rather than raw accuracy (Dupple).

Best for: Businesses translating formatted documents at scale who need glossaries, formality and data controls.

6. Qwen3.7-Max and Qwen-MT — best for Asian languages

Price: Qwen3.7-Max API $2.50/$7.50 per million tokens; Qwen app free; open Qwen weights self-hostable Languages: 119+ in the app; 201 in the open models; Qwen-MT dedicated API Best for: Chinese, Japanese, Korean and other Asian languages; low-cost multilingual work

Alibaba’s Qwen3.7-Max is the strongest system for Asian languages, handling Chinese idioms, Japanese honorifics (keigo) and Korean formal speech levels more naturally than Western-trained models (RoundTalk). The dedicated Qwen-MT translation API adds glossary control and formality settings and maintains high terminology accuracy for technical and legal documentation (Machine Translate). The free Qwen app covers 119+ languages; the open-weight Qwen models support 201 and can be self-hosted.

Limitations: Qwen is built by a China-headquartered company, so for sensitive or regulated work the safer route is self-hosting the open weights rather than using the hosted app.

Best for: Anyone whose work centres on Chinese, Japanese, Korean or other Asian languages, and cost-sensitive multilingual teams.

7. DeepSeek V4 — best for cost and self-hosting

Price: API $0.14/$0.28 per million tokens; MIT licence (self-hostable free) Best for: High-volume translation on a budget, air-gapped or data-sovereign pipelines

DeepSeek V4 is the value and privacy play. Its MIT licence means you can run a capable multilingual model entirely on your own infrastructure with no data leaving your network, and its hosted API is the cheapest of any frontier-class model at $0.14/$0.28 per million tokens — roughly a hundredth of a flagship LLM. It is particularly strong on Chinese–English.

Limitations: Serving quality depends on your stack, and it trails the top proprietary models on the hardest, most nuanced pairs.

Best for: Developers translating at very high volume, and teams that need translation to stay on-premises.

8. Consumer and specialist apps

• Papago (Naver) — the best free app for Korean, Japanese, Chinese and Southeast Asian languages, with contextual handling of honorifics, keigo and measure words that general apps often flatten (Unstar). Free with limited ads.
• iTranslate — subscription all-in-one covering 100+ languages; Pro at $5.99/month or $39.99/year unlocks offline packs, AR camera translation, voice-to-voice conversation, verb conjugations and an Apple Watch app.
• Microsoft Translator — 100+ languages, deep Office, Teams and Edge integration, and the cheapest major API (below). Free apps for phone, conversation and camera translation.

9. Localisation platforms — best for websites, apps and software

When translation has to feed continuously into a product, a translation management system (TMS) wraps AI translation in workflow, review, glossaries and developer integrations:

• Weglot — best no-developer website translation: automatic content detection on any CMS, multilingual SEO, a free tier, and five-minute setup on WordPress (withallo).
• Lokalise — best for product and engineering teams: an AI orchestration layer that evaluates multiple engines per string and keeps localised releases in sync with the source. From $144/month (Explorer); AI translation on the $499/month Growth plan (Ringly).
• Crowdin — best for developer-led teams: an AI stack drawing on OpenAI, Anthropic and Azure, with 700+ integrations and connectors.
• Smartling — best for enterprises with dedicated localisation staff: AI translation plus human validation, continuous localisation and GitHub integration.
• Phrase — best for governance: segment-level translation history, glossary linkage and audit-friendly, traceable records for regulated work.

Feature comparison: the full matrix

Use-case specific recommendations

For free everyday translation

Winner: Google Translate (Gemini 3.5) — 249 languages, Gemini-quality text, live speech and camera modes, all free. Alternative: the free tiers of ChatGPT or Gemini when you want to steer the tone.

For professional, nuanced text

Winner: a frontier chat model — GPT-5.6 Sol for creative and marketing copy, Claude Opus 4.8 for long-document consistency. Both let you specify register, audience and glossary. Alternative: DeepL if formatting preservation matters more than nuance.

For business documents with formatting

Winner: DeepL — preserves Word, PowerPoint and PDF layout, with grammatically-aware glossaries and formality control. Alternative: a localisation platform for continuous document pipelines.

For Chinese, Japanese and Korean

Winner: Qwen3.7-Max or Qwen-MT — the most reliable on honorifics, keigo and measure words. Alternative: Papago for free consumer use.

For lowest cost at scale

Winner: Microsoft Translator at $10/M characters (2M free monthly), or DeepSeek V4 at $0.14/$0.28 per million tokens for self-hosting. Alternative: Google Cloud Translation at $20/M characters.

For websites and apps

Winner: Weglot for no-code sites; Lokalise or Crowdin for product teams that need continuous localisation and developer integrations. Alternative: Smartling for enterprise with human review.

For real-time spoken translation

Winner: Gemini 3.5 Live Translate — natural speech-to-speech in 70+ languages inside Google Translate. Alternative: the voice modes in ChatGPT or the Gemini app for conversational back-and-forth.

For maximum privacy and data control

Winner: self-hosted DeepSeek V4 (MIT) or open Qwen weights — capable multilingual translation with no data leaving your infrastructure. Alternative: DeepL’s enterprise data controls for a hosted option.

Pricing comparison: what you’ll actually pay

Dedicated translation APIs (per million characters, USD)

LLM APIs (per million tokens, USD)

How to read this: character-priced engines and token-priced LLMs are not directly comparable. One million characters of English is roughly 250,000 tokens, and LLMs bill for output as well as input, so translating a million characters through an LLM API typically costs several times a character-priced NMT engine (SimpleLocalize). Above about 10 million characters a month, a character-priced engine is usually the more predictable and cheaper choice; below it, an LLM’s nuance and steerability often justify the premium (Langbly).

Consumer plans

Cost strategy: for free everyday use, Google Translate is unbeatable. For paid work, route by task — a cheap NMT engine or open model for bulk and formulaic strings, a frontier LLM for the nuanced 10–20%, and a localisation platform when translation has to run continuously into a product.

What users and researchers think

The “LLMs won on quality” consensus is now mainstream

The WMT25 findings and multiple independent 2026 benchmarks agree that general-purpose LLMs have caught or passed dedicated engines on human-rated quality for high-resource pairs (intlpull). The caveat researchers repeat: this holds for well-resourced languages and everyday domains, not for low-resource languages or highly specialised material, where dedicated systems and human post-editing still matter.

Metric literacy is the new dividing line

The recurring lesson is that BLEU and COMET disagree, and buyers who quote a single leaderboard number without naming the metric get burned. COMET and human evaluation favour LLMs; BLEU favours engines on short strings. Serious teams now evaluate on their own content, in their own languages, rather than trusting a headline figure.

Workflow beats raw accuracy for businesses

For teams shipping products, the deciding factors are formatting preservation, glossary control, data governance and integration — which is why DeepL, Qwen-MT and the localisation platforms retain strong loyalty even as raw-quality gaps close. A slightly less nuanced translation that lands correctly formatted, on-brand and in the codebase often beats a marginally better sentence that needs manual cleanup.

Asian-language users prefer regional tools

Consistent community sentiment holds that Qwen and Papago handle Korean honorifics, Japanese keigo and Chinese measure words better than Western-built general systems (Unstar) — a reminder that “best translator” is language-pair-specific, not global.

Recent launches reshaping the market (2026)

Gemini 3.5 Live Translate (June 2026). Google shipped real-time speech-to-speech translation in 70+ languages that preserves the speaker’s intonation, pacing and pitch, and began rolling it into Google Translate (WinBuzzer).

Gemini-powered Google Translate text upgrade (2026). Google rebuilt Translate’s text quality on Gemini for more natural, idiom-aware output, rolling out first in the US and India across English and nearly 20 languages (Google).

DeepL next-gen language model and agentic tools. DeepL moved its engine to an LLM-based next-gen model — letting Pro users switch between next-gen and classic per job — and expanded into agentic productivity tooling (DeepL).

WMT25 findings (November 2025). The field’s flagship shared task moved to harder, document-level test sets and evaluated LLMs alongside dedicated providers, formalising the LLM-versus-engine contest under professional human evaluation (ACL Anthology).

Localisation cost collapse (2025–2026). AI-assisted localisation fell from roughly $0.20 to $0.002 per word — a 100x drop — making machine-first localisation the default and reshaping the TMS market around orchestration and review rather than raw translation (withallo).

Frequently asked questions

What is the best AI for translation in 2026?

For free everyday use, Google Translate, now built on Gemini 3.5 and covering 249 languages. For nuanced professional text, a frontier chat model — GPT-5.6 Sol, Claude Opus 4.8 or Gemini 3.5 Pro — now matches or beats dedicated engines on human evaluation for high-resource languages. For formatted business documents, DeepL; for Asian languages, Qwen3.7-Max.

Is ChatGPT or Google Translate better for translation?

They serve different needs. Google Translate is faster, free and covers 249 languages — best for quick, everyday translation. ChatGPT is more accurate on nuanced, context-dependent text because you can instruct it on tone, audience and terminology, and it adapts idioms rather than translating them literally. For a one-off phrase use Google Translate; for a nuanced paragraph use ChatGPT.

Is DeepL still better than ChatGPT and Google?

Not clearly, anymore. DeepL’s own blind tests claim it beats GPT-4, Google and Microsoft (DeepL), but independent 2026 human-evaluation round-ups place frontier LLMs at or above DeepL for high-resource pairs. DeepL’s durable advantages are document formatting, grammatically-aware glossaries, formality control and predictable per-character pricing — workflow, not raw accuracy.

What is the most accurate AI translator?

It depends on the metric. On COMET and human evaluation (which track meaning and nuance), frontier LLMs lead. On BLEU (surface overlap, best for short strings), dedicated engines like DeepL still win. For Chinese, Japanese and Korean, Qwen is the most reliable. There is no single accuracy champion across all languages and content types.

What is the cheapest AI translation API?

Microsoft Translator at $10 per million characters, with 2 million characters free every month — the cheapest major dedicated API. For LLM-based translation, DeepSeek V4 at $0.14/$0.28 per million tokens (and free to self-host under MIT) is the floor. Google Cloud Translation is $20/M characters; DeepL is $25/M characters.

Can AI replace human translators?

For most everyday and business content, AI now reaches an estimated 85–90% of professional-translator quality at 5–10% of the cost (GetBlend), so it handles the bulk of routine translation. Humans remain essential for high-stakes, legal, literary, marketing-critical and low-resource-language work, where the standard workflow is now AI translation plus human post-editing rather than translation from scratch.

What is the best AI for translating Asian languages?

Qwen3.7-Max and the dedicated Qwen-MT API for professional and API use — they handle Chinese idioms, Japanese keigo and Korean honorifics most reliably (RoundTalk). For free consumer use, Papago (Naver) is the strongest app for Korean, Japanese and Chinese.

What is the best AI to translate a website or app?

A localisation platform, not a chat model. Weglot is the best no-developer option for websites (any CMS, multilingual SEO, free tier); Lokalise and Crowdin suit product and engineering teams needing continuous localisation and developer integrations; Smartling fits enterprises that want AI translation with human validation.

Which AI translates documents while keeping the formatting?

DeepL is the strongest for this — its document translation preserves Word, PowerPoint and PDF layout and supports grammatically-aware glossaries and formality control. Google Translate and localisation platforms also handle document formatting; raw chat models like ChatGPT and Claude generally do not preserve complex layouts automatically.

Conclusion: how to choose in July 2026

Translation is the use case where the LLM wave has most quietly changed the answer. The dedicated engines that defined machine translation for a decade are now, on human-rated quality for major languages, matched or beaten by general-purpose models — while keeping real advantages in formatting, glossaries and cost.

• Best free everyday translator: Google Translate (Gemini 3.5), 249 languages.
• Best for nuanced professional text: a frontier chat model — GPT-5.6 Sol or Claude Opus 4.8.
• Best for formatted business documents: DeepL.
• Best for Asian languages: Qwen3.7-Max / Qwen-MT, or Papago for free.
• Cheapest at scale: Microsoft Translator ($10/M chars) or self-hosted DeepSeek V4.
• Best for websites and apps: Weglot, Lokalise or Crowdin.
• Best for real-time speech: Gemini 3.5 Live Translate.

Pick by the job: the free app for daily life, the steerable chat model for nuance, the dedicated engine for formatted documents, the regional specialist for Asian languages, and a localisation platform when translation has to run continuously into your product. And whatever the leaderboard says, evaluate on your own content — because in translation, “best” is always a specific language pair, a specific metric and a specific kind of text.

For the underlying models behind these tools, see the best AI models ranking; for related workflows, see best AI for writing and best AI for transcription.

Quality varies by language pair, metric and content type; where vendor and independent numbers differ, we cite both and flag which is which. Pricing and availability change frequently — verify against official sources before relying on a figure.