Best AI Translation Tools 2026 Ranked

Translation looks like a solved problem from the outside. You paste English in, Spanish comes out, everyone is happy. Then your localization team tells you the marketing copy reads like a terms-of-service document, your medical records platform is misgendering patients, or the Japanese output for your e-commerce site is technically correct but sounds like it was written by someone who has never bought anything online.

This comparison covers the full stack: raw API engines (DeepL, Google, Microsoft, Amazon), frontier LLMs used for translation (GPT-5, Claude 4, Gemini 2.5 Pro), open-weight alternatives (Qwen3, DeepSeek-V3), enterprise translation management systems with AI layers (Lokalise, Phrase, Crowdin, Smartling, Transifex, Smartcat), hybrid human-AI workflows (Unbabel), and developer-friendly open-source options. I also flag the live meeting translation tools and consumer devices for completeness, though those are a different product category.

For the benchmark data behind these rankings, see the Machine Translation Benchmarks Leaderboard 2026.

Quick Rankings

COMET scores from IntlPull 2024 benchmark and WMT studies. BLEU is English-to-German. LLM pricing is per-token from official rate cards at time of writing. "Depends on engine" - TMS tools plug into your choice of MT backend.

Methodology

I'm using three layers of signal to rank these tools.

Automatic metrics are the starting point. BLEU measures n-gram overlap between output and reference translation - it's a legacy metric, useful for tracking trends and comparing against historical baselines, but it punishes paraphrase and doesn't handle synonyms well. Anything over 60 for European pairs is generally solid; use it as a baseline, not a verdict. chrF++ uses character-level n-gram F-score and correlates better with human judgment on morphologically rich languages like Finnish, Czech, or Turkish. COMET is a neural metric trained on human quality judgments and scores roughly 0.75-0.95 for production systems - it's the most reliable automatic metric currently available, but the WMT24 finding that TOWER-v2-70B gamed it (ranking first on COMET across all 11 language pairs while losing to Claude 3.5 on 9 of those same pairs under human evaluation) is a real warning.

Human evaluation is the gold standard. WMT's Error Span Annotation (ESA) process has professional human annotators mark every error span in translation outputs. This is expensive, which is why it happens once per year on final systems. WMT24 and WMT25 are the two most recent datasets and are the most trustworthy signal we have for frontier systems.

Structural factors - I also factor in pricing, language coverage, API design quality, latency, streaming capability, and whether the tool's marketing claims hold up under scrutiny. "Translation LLM" that turns out to be standard NMT gets flagged. Enterprise pricing for a GPT-4 wrapper also gets flagged.

Frontier LLMs for Translation

The most important thing to understand about the 2026 translation landscape: WMT25 human evaluation placed LLMs in the top four positions, with dedicated MT systems like Google Translate, DeepL, and Microsoft Translator ranked mid-table. That's a significant shift from even 2023, when dedicated MT held its own against LLMs.

Gemini 2.5 Pro

Gemini 2.5 Pro won WMT25, Google's own annual shared task, by topping 14 of 16 tested language pairs on human evaluation. That is the highest quality bar publicly demonstrated by any system as of April 2026.

The caveat is cost. At $1.25/1M input tokens and $10/1M output tokens (1M+ context tier), Gemini 2.5 Pro is not a viable replacement for Google Translate when you're localizing 100,000 strings for an e-commerce SKU catalog. For content where quality genuinely matters - marketing campaigns, legal documents, high-visibility product pages - the quality differential can justify the price.

What it does: State-of-the-art neural translation via the Gemini API. No dedicated translation endpoint; you call the chat completions API with a translation system prompt. Supports 100+ languages with strong multilingual training data.

Pricing: $1.25/1M input, $10/1M output (1M+ context). Roughly $0.02-0.04 per 1,000 translated words at typical compression ratios.

Best fit: Premium quality localization for marketing, legal, or brand-voice-sensitive content. Also the right default when your target language pair is non-European and LLM multilingual corpora outperform specialized NMT.

Honest gotcha: You're calling a chat API, not a translation engine. Prompt engineering matters for consistency, glossary adherence, and tone. For high-volume deterministic localization, the lack of a specialized translation endpoint is a real gap.

GPT-5 / GPT-4.1

GPT-4.1 ranked second at WMT25, close behind Gemini 2.5 Pro. The IntlPull 2024 benchmark put GPT-4 Turbo at 0.847 COMET, ahead of DeepL API Pro (0.838) - and GPT-5 / 4.1 score higher still. For East Asian languages, GPT-4o already outperforms DeepL on BLEU (57.4 vs 51.3 for English-to-Chinese, 54.8 vs 49.1 for English-to-Japanese), and newer models extend that lead.

What it does: General-purpose LLM API for translation. No dedicated MT endpoint. Strong for complex content, stylistic adaptation, and language pairs with large multilingual training data.

Pricing: GPT-5 pricing was set at $2/1M input, $8/1M output at launch. GPT-4.1 is $2/1M input, $8/1M output (standard). GPT-4.1 mini is available at lower cost for lighter work.

Best fit: Teams already using OpenAI for other tasks who want a single vendor. Strong for English-to-Asian language pairs and for content requiring tone or register adaptation.

Honest gotcha: Same caveat as Gemini 2.5 Pro - you're calling a chat API. Consistency across large glossaries requires explicit prompt engineering or function calling for term lookup. At GPT-5 pricing, high-volume localization is expensive.

Claude Opus 4.x and Claude 4

Claude 3.5 Sonnet placed first on WMT24 human ESA scores in 9 of 11 language pairs - the 2024 equivalent of Gemini 2.5 Pro's 2025 result. Claude 4 and Opus 4.x extend that position. The Localize.js blind study put Claude 3.5 Sonnet at 4.92/5 for Chinese and 4.82/5 for Japanese - the two highest single-language scores in that study.

Where Claude stands out is stylistic fluency. Human evaluators consistently rate Claude outputs as more natural-sounding than equivalently-scored alternatives on automatic metrics. For marketing copy, creative content, or any translation where voice matters, that's worth knowing.

What it does: LLM API translation with notably strong stylistic output and high human evaluation scores for East Asian languages.

Pricing: Claude Opus 4 at $15/1M input, $75/1M output. Claude Sonnet 4 at $3/1M input, $15/1M output. Sonnet 4 is the more practical choice for most translation workloads.

Best fit: Creative localization, marketing copy, content where tone and voice need to carry across languages. Also strong for Chinese and Japanese.

Honest gotcha: Opus 4 is expensive for volume work. Use Sonnet 4 unless you have a specific quality requirement that the benchmark gap justifies. See the Translation Benchmarks Leaderboard for the full metric comparison.

Dedicated MT Engines

DeepL API Pro

DeepL remains the strongest dedicated MT engine for European language pairs as of April 2026. The IntlPull benchmark puts it at 0.838 COMET - slightly behind Claude 3.5 Sonnet (0.841) and GPT-4 Turbo (0.847), but ahead of Google and Microsoft dedicated engines. The BLEU gap for European pairs is real: 64.5 for English-to-German vs. 61.2 for GPT-4o.

The practical advantages of DeepL over frontier LLMs for high-volume work are determinism, speed, and cost. You get consistent output on repeated inputs. Latency for a single-sentence API call is under 200ms. At the volume tiers where enterprise localization operates - millions of characters per month - DeepL's per-character pricing is materially lower than LLM token pricing.

DeepL also offers DeepL Write for grammar and style correction, and a glossary API for enforcing terminology consistency across large translation batches. The glossary feature is genuinely useful for technical or legal content where term consistency matters more than paraphrase quality.

What it does: Dedicated neural MT engine covering 33 language pairs. Translation API with glossary support, formal/informal register selection, and real-time streaming.

Pricing: Starter plan $5.99/month (500K chars/month). Advanced $29.99/month (1M chars). Ultra $65.99/month (5M chars). Business $119.99/month (10M chars). Enterprise custom. API-only access via Pro API subscription.

Best fit: High-volume EU language localization. Technical documentation, legal text, software UI strings. Any workflow where consistent and predictable output matters more than peak quality.

Honest gotcha: DeepL's Chinese and Japanese performance lags frontier LLMs by a meaningful margin (BLEU 51.3 vs GPT-4o's 57.4 for EN-ZH). If East Asian languages are a primary use case, DeepL is not your best option at any price point. Coverage is also limited to 33 language pairs - well below Google or Microsoft.

Google Cloud Translation API

Google offers three translation tiers and calling them all "Translation API" obscures meaningful differences.

Basic (V2) is the underlying infrastructure most users hit from the web interface. It's a mature NMT system - solid for high-resource European pairs, competitive for major Asian languages, and meaningfully cheaper than other options at scale.

Advanced (V3) adds glossaries, batch translation, model selection, and per-project data isolation. This is the right API tier for enterprise use cases. Pricing at $20/1M characters for Advanced is higher than the Basic tier but competitive with Microsoft.

Translation LLM is Google's highest tier, added in 2024 and upgraded in 2025, and I want to be direct about what it is: it's Google's standard NMT stack with a language model adaptation layer added for context-aware refinement, marketed with LLM branding. The quality is better than Basic V2 on complex sentences, but positioning it as a distinct "LLM translation" product implies a fundamental architectural shift that the benchmark data doesn't support. In WMT25, Google Translate sits mid-table behind the actual LLMs and behind DeepL for European pairs. The Translation LLM tier may or may not be worth the premium depending on your content type - test it on your actual content before paying for it.

What it does: 135-language neural MT with cloud-native integration (GCP pub/sub, Cloud Storage, BigQuery). Batch and real-time endpoints. Glossary API. Basic V2 is the best-value tier for GCP-native workflows.

Pricing: Basic V2 - $20/1M chars. Advanced V3 - $20/1M chars (same rate, different features). Translation LLM - higher pricing at custom/enterprise tier.

Best fit: Deep GCP integrations where switching vendor costs exceed the per-character savings. Broad language coverage at scale (135 languages vs DeepL's 33). AutoML model fine-tuning for domain-specific content.

Honest gotcha: If you're not already in GCP, the integration overhead rarely justifies the cost vs DeepL or Microsoft for European pairs. The "Translation LLM" marketing overstates the product differentiation.

Microsoft Translator API

Microsoft Translator covers 135+ languages and integrates natively with Azure Cognitive Services. The benchmark numbers (roughly 0.81 COMET, around 61 BLEU for EN-DE) place it solidly below DeepL for European pairs and below frontier LLMs for everything, but the Azure integration story is real for enterprise customers already on Microsoft infrastructure.

The Real-Time Translator feature uses WebSocket-based streaming for conversational translation, useful for live meeting captioning pipelines. Microsoft has also added Custom Translator - a fine-tuning layer for training domain-specific MT models on your content. Custom models can close the gap with DeepL for specialized domains where you have sufficient parallel training data.

What it does: 135-language MT API with real-time streaming, Custom Translator fine-tuning, and Azure Active Directory integration. Transliteration supported for script conversion.

Pricing: $10/1M characters (S1 standard tier). Custom Translator training: $10/1M training characters. First 2M characters/month free under the S0 tier.

Best fit: Azure-native enterprise deployments. Real-time translation pipelines via WebSocket streaming. Teams with sufficient domain-parallel data to justify Custom Translator fine-tuning.

Honest gotcha: Benchmark quality trails DeepL for European pairs and trails LLMs for complex content. The Azure discount you're getting by staying on-platform needs to outweigh the quality gap.

Amazon Translate

Amazon Translate is the MT option for AWS-native pipelines. Pricing at $15/1M characters positions it between Microsoft ($10) and Google Advanced ($20), but the benchmark profile - roughly 0.80 COMET, human quality score 4.62/5 in the Localize.js study - puts it below Google and Microsoft on automatic metrics.

Custom Terminology and Active Custom Translation let you inject terminology lists and fine-tune on your content, similar to Microsoft's Custom Translator. The AWS-native event pipeline integration (Lambda, S3, EventBridge) is the real value proposition.

What it does: Neural MT for 75 languages with AWS IAM auth, S3 batch processing, and real-time API. Custom Terminology for glossary enforcement.

Pricing: $15/1M characters standard. Custom translation additional. First 2M characters/month free for 12 months (new accounts).

Best fit: AWS-native data processing pipelines. Teams running content through Lambda or SageMaker who need translation as a component, not a primary product.

Honest gotcha: Benchmark quality is behind Google, Microsoft, and DeepL at a higher per-character price. It's primarily the right choice because of AWS integration, not because of translation quality.

Open-Weight LLMs for Translation

DeepSeek-V3 and Qwen3

These two open-weight LLMs deserve specific mention for translation workloads. Both ranked in WMT25 human evaluation (DeepSeek-V3 at rank 4, Qwen3 at rank 7), placing above legacy dedicated MT engines like Google Translate, DeepL, and Microsoft Translator in that evaluation.

DeepSeek-V3 is available via API at $0.27/1M input tokens - roughly 10x cheaper than GPT-4.1 at equivalent parameter scale. For teams running high-volume translation where LLM quality is worth the complexity but frontier pricing is prohibitive, DeepSeek-V3 via the official API is a viable option. Self-hosting on your own hardware adds data privacy for sensitive content.

Qwen3 (72B) is particularly strong for Chinese-centric workflows. Given Alibaba's Chinese training data scale, Qwen3 outperforms several frontier models for Chinese translation pairs. It's open-weight and available for self-hosting, which matters for legal or healthcare workflows where data cannot leave your environment.

Honest gotcha: Both tools require prompt engineering discipline - they're chat APIs, not translation engines. Consistency at scale needs careful system prompt design. Self-hosting at production throughput requires meaningful GPU infrastructure.

Enterprise TMS Platforms with AI

Enterprise Translation Management Systems (TMS) are the workflow layer above raw MT engines. They handle glossaries, translation memories, review queues, project management, and integrations with your content pipeline. The AI features in these platforms are mostly one of two things: an MT engine connector (you pick DeepL, Google, OpenAI, or their own engine), or LLM-powered post-editing and quality estimation.

The key thing to understand: the ceiling quality of a TMS is determined by which MT engine you connect, not by the TMS itself. If a vendor is charging enterprise prices and their "AI translation" is sitting on top of Google Basic V2, you're paying for workflow features, not translation quality. Ask the vendor explicitly which MT models power their AI translation.

Lokalise AI

Lokalise is developer-first by design. The workflow is built around CI/CD - you push strings from GitHub or GitLab, Lokalise translates them, and you pull reviewed translations back on merge. The SDK and CLI are solid. Translation memory and glossary enforcement work well out of the box.

The AI translation layer supports DeepL, Google, Microsoft, and OpenAI as MT backends, plus their own Lokalise AI engine. The quality is directly tied to whichever backend you configure.

Pricing: Start plan $168/month (annual), 10 seats included. Essential from $352/month. Enterprise custom. All plans include unlimited words at scale.

Best fit: Software localization for developer teams. GitHub/GitLab integrated workflows. Startups to mid-market SaaS companies.

Honest gotcha: Pricing is per seat, not per word - it scales well for large teams but is expensive for small teams with high translation volume.

Phrase

Phrase (formerly Memsource) is a mature enterprise TMS with the broadest feature set in this category. Translation memory, a computer-assisted translation (CAT) tool, QA automation, and a connector to virtually every MT engine available. The Phrase AI feature set includes LLM-powered pre-translation, segment-level quality estimation, and AI-assisted post-editing.

The platform is designed for localization teams at large enterprises and translation agencies. The workflow sophistication exceeds what developer-focused tools like Lokalise offer, at corresponding complexity and price.

Pricing: Custom enterprise. Starter tiers exist for smaller teams; contact sales for enterprise contracts.

Best fit: Large enterprise localization teams, translation agencies managing multi-client projects, content-heavy organizations with complex review workflows.

Honest gotcha: Onboarding complexity is high. For a small team or a developer trying to localize a single app, Phrase is overkill. The AI features are only as good as the MT engine you plug in.

Crowdin AI

Crowdin is the TMS that open-source projects reach for first - it's free for public open-source projects, which is genuinely useful. The AI features include pre-translation via GPT-4o, DeepL, or other configured engines, plus AI-powered QA and glossary suggestions.

The in-context editor for software localization - where translators see strings in the actual UI context - is one of the better implementations in this space.

Pricing: Basic plan $50/month (annual). Standard $99/month. Business $199/month. Enterprise custom. Open-source: free for public repos.

Best fit: Open-source software localization. SaaS products with developer-centric localization needs. Teams that want strong in-context translation tooling.

Honest gotcha: Feature depth trails Phrase for complex enterprise workflows. AI quality depends entirely on which backend you configure.

Smartling AI

Smartling's pitch is brand-consistent localization at scale. The platform uses translation memories, style guides, and AI glossary enforcement to keep output on-brand across millions of words. The review workflow is structured for enterprise compliance.

Smartling also operates their own linguist network for hybrid workflows - you can route AI-translated content to human review in the same platform, with quality scores driving review prioritization.

Pricing: Custom enterprise. No self-serve pricing published.

Best fit: Enterprise marketing teams and global brands that need translation to maintain consistent voice and tone across markets. Regulated industries requiring human review at scale.

Honest gotcha: Pricing opacity is a red flag for smaller buyers. No self-serve option.

Transifex AI

Transifex targets software and content teams with a reasonably priced self-serve tier. The AI layer uses LLMs (OpenAI-based) for pre-translation and QA suggestions. Translation memory is solid. The GitHub and Figma integrations are useful for design and dev teams.

Pricing: Starter $25/month (annual). Growth $170/month. Scale $450/month. Enterprise custom.

Best fit: Mid-market software and content teams who need more than a basic file-upload workflow but don't need Phrase's enterprise complexity.

Honest gotcha: AI quality depends on which backend is configured. The OpenAI-powered default is GPT-4o-class quality, which is competitive but not best-in-class.

Smartcat AI

Smartcat's differentiator is a built-in freelance translator marketplace alongside the TMS. You can route content directly to vetted human translators from the same platform where AI pre-translates it. The business model includes a marketplace fee rather than a pure SaaS subscription, which can be cost-effective for variable translation volume.

Pricing: Free tier available with limited volume. Paid plans from custom pricing. Transaction fees on marketplace work.

Best fit: Teams who need both AI and human translation in a single workflow. Variable volume workloads where per-project pricing is preferable to flat subscription.

Honest gotcha: The marketplace model adds coordination overhead for teams that want a pure SaaS workflow. AI quality depends on your configured MT backend.

Unbabel - Human Plus AI

Unbabel is the serious option when quality cannot be left to machines alone. Their pipeline uses MT as a first pass, followed by trained human editors who post-edit the output, with quality scoring throughout. The model is positioned at the enterprise end of the market - think global customer support operations, financial services, and e-commerce at significant scale.

The TOWER-v2 translation model - which won COMET at WMT24 before human evaluation exposed the metric gaming - is Unbabel's own research output. They're not just a vendor; they're active MT researchers.

What it does: MT + human post-editing workflow at scale. Quality tiers from pure MT to full human review. SLA-backed delivery with measurable quality scores.

Pricing: Custom enterprise. Not self-serve.

Best fit: Enterprise customer support localization. Legal, medical, or financial content where MT errors carry real liability. Any use case where machine-only output is not acceptable for compliance or quality reasons.

Honest gotcha: Cost and lead time are both higher than pure MT. For content that doesn't require human oversight, you're paying for something you don't need. The WMT24 COMET gaming incident is public knowledge - it's worth knowing that TOWER-v2 was optimized for the metric, not for underlying quality. The human edit layer corrects for this, but it's a data point worth having.

Open-Source NMT: NLLB-200, OpenNMT, Marian

If your use case involves low-resource languages, domain-specialized content, or strict data privacy requirements, the open-source NMT ecosystem is worth knowing.

NLLB-200 (Meta AI, No Language Left Behind) covers 200+ languages with models from 600M to 54B parameters. The 3.3B variant fine-tuned on medical data beats one-shot Llama 3.1 405B by 10+ BLEU points on TICO-19 English-to-French medical content. For Swahili, Yoruba, Bhojpuri, and other low-resource languages where frontier LLMs have limited training data, NLLB-200 is often the most reliable option.

OpenNMT is the reference implementation for neural MT research. Production use is possible but requires engineering work. More relevant as a training framework for custom models than as a production API.

Marian NMT (now wrapped by the Helsinki NLP OpusMT project) provides pre-trained models for hundreds of language pairs, available via the Hugging Face hub. Many of these are specialized for news, legal, or medical domains. For researchers and teams wanting to fine-tune on domain data, Marian is a practical starting point.

Honest gotcha: Self-hosting at production throughput requires GPU infrastructure. A single A100 can handle reasonable translation throughput for NLLB-200 3.3B, but capacity planning matters. There's no managed API - you build the pipeline.

Live Meeting Translation and Consumer Devices

For completeness: this is a distinct product category from the translation APIs and TMS platforms above, and the evaluation criteria are different.

LanguageLine and Wordly focus on real-time spoken language interpretation - conference calls, live events, meetings. Wordly uses AI for real-time simultaneous interpretation; LanguageLine uses professional human interpreters for legal and medical contexts where accuracy has direct consequences.

Timekettle and Pocketalk are hardware translation earbuds and devices targeting consumer use cases: travel, international conversations, casual multilingual meetings. They use cloud MT APIs under the hood and are not relevant for enterprise localization pipelines.

Best Pick by Use Case

Highest quality, budget is secondary - Gemini 2.5 Pro for WMT25-verified human evaluation performance across 14 of 16 language pairs. Use the Gemini API directly with a structured translation system prompt.

East Asian languages (Chinese, Japanese, Korean) - GPT-4.1 or Claude Sonnet 4 for quality. DeepSeek-V3 or Qwen3 if cost matters and self-hosting is acceptable.

High-volume EU language localization - DeepL API Pro. The BLEU advantage for European pairs is documented, the per-character pricing is competitive, and the output is deterministic. Reserve LLMs for content where voice matters.

GCP or Azure or AWS-native - Use your cloud provider's MT API for the integration simplicity, then evaluate whether the quality delta from DeepL or a frontier LLM is worth switching vendors for your specific language pairs.

Enterprise software localization - Lokalise for developer-team workflows. Phrase for large enterprise localization programs. Crowdin if you're an open-source project.

Medical, legal, or regulated content - Fine-tune NLLB-200 on your domain data for low-resource languages. Unbabel for enterprise-scale workflows where human review is non-negotiable. Verify terminology consistency independently of COMET scores - generic MT metrics don't measure what matters here.

Maximum data privacy - Self-hosted Qwen3 or NLLB-200. Neither leaves your network.

FAQ

Which AI translation tool is most accurate in 2026?

Gemini 2.5 Pro won WMT25 human evaluation, leading 14 of 16 tested language pairs. GPT-4.1 placed second. For European language pairs specifically, DeepL API Pro holds an edge on BLEU (64.5 EN-DE vs 61.2 for GPT-4o). Human evaluation and automatic metrics sometimes pick different winners - the Translation Benchmarks Leaderboard has the full breakdown.

Is DeepL still the best MT API?

Best for high-volume European language pairs, yes. The BLEU advantage over cloud MT APIs is documented and the deterministic output makes quality assurance easier at scale. For non-European language pairs, frontier LLMs (GPT-4.1, Gemini 2.5 Pro, Claude 4) now outperform DeepL on human evaluation.

What is Google's "Translation LLM" and is it worth the extra cost?

Google's Translation LLM tier adds a language model refinement layer on top of their existing NMT pipeline. Benchmark results don't show a step-change in quality that justifies calling it a distinct "LLM" product versus an improved NMT system. In WMT25, Google Translate sits mid-table behind actual frontier LLMs. Whether the Translation LLM tier is worth the premium over Advanced V3 depends on your content type - test on your actual data before committing.

Should I use a TMS platform or call a translation API directly?

Directly for developer use cases where you're processing content programmatically - the TMS layer adds cost without value if you don't need workflow management, human review routing, or translation memory. TMS platforms earn their value for teams with human translators in the loop, multi-language simultaneous projects, glossary enforcement across thousands of strings, or compliance requirements around review and sign-off.

Can open-weight LLMs (Qwen3, DeepSeek-V3) match frontier APIs for translation?

For general content in high-resource language pairs, yes at competitive quality levels. DeepSeek-V3 ranked 4th in WMT25 human evaluation. Qwen3 is particularly competitive for Chinese pairs. The tradeoff is infrastructure overhead for self-hosting versus using the official APIs, and prompt engineering discipline to get consistent output at scale.

What is COMET and why does it matter?

COMET is a neural evaluation metric trained on human translation quality judgments. It correlates with human assessment better than BLEU, and scores typically range 0.75-0.95 for production systems. The caveat: WMT24 found that TOWER-v2-70B gamed COMET by optimizing against the metric, ranking first on COMET while losing to Claude 3.5 on 9 of 11 language pairs under human evaluation. Use COMET as a development signal; use ESA human evaluation for final production decisions.

Sources

• WMT25 General MT Shared Task findings - Human evaluation results, Gemini 2.5 Pro wins 14/16 language pairs
• WMT24 General MT Shared Task findings - Claude 3.5 Sonnet human ESA scores, TOWER-v2 COMET gaming
• WMT25 Preliminary Results - Slator - Coverage of WMT25 system rankings
• IntlPull LLM Translation Benchmark 2026 - COMET scores by provider
• AI Translation Blind Study - Localize.js - Human ratings across Chinese, Japanese, Spanish, German, French
• Domain-Specific LLM Translation (arXiv 2412.05862) - NLLB-200 vs LLMs on TICO-19 medical content
• DeepL LLM Translation Quality Claims - DeepL's own benchmark methodology
• Cloud Translation Pricing - Google - Basic V2 and Advanced V3 per-character rates
• Amazon Translate Pricing - Standard MT per-character rates
• Azure AI Translator - Microsoft - S1 pricing tier
• Translation Model Benchmark - Lara Translate (Feb 2026) - Comparative benchmark data
• Pitfalls and Outlooks in Using COMET (WMT24) - COMET metric gaming analysis