Best translation APIs in 2026: NMT engines, LLMs, and how to choose

The translation API landscape changed substantially between 2024 and 2026. The decision that used to be "which NMT engine should I integrate" has become a two-layer question: do you need an NMT engine (fast, cheap, predictable) or an LLM-based API (higher quality ceiling, more expensive, slower), and which specific option within that category fits your use case?

According to Intento's State of Translation Automation 2025, LLMs now represent 89% of top performers across language pair evaluations, up from 55% the year before. For most quality-critical translation workflows, LLM APIs have overtaken NMT engines on output quality. For latency-sensitive or high-volume workflows where cost per character matters, NMT engines still hold a significant advantage.

This guide covers both categories, with current pricing, honest trade-offs, and a practical framework for deciding which fits your needs.

In this article

1. NMT translation APIs: the main options
2. LLM translation APIs: what changes with this category
3. How NMT and LLM APIs compare on quality, speed, and cost
4. API pricing comparison table (2026)
5. How to choose the right API for your use case
6. A note on multi-engine approaches
7. FAQs

NMT translation APIs: the main options

Traditional NMT (Neural Machine Translation) APIs are still the foundation of most translation integrations. They are fast, priced by the character, predictable in latency, and mature in terms of documentation and ecosystem support.

Google Cloud Translation API

Google Cloud Translation is one of the most widely used translation APIs. It supports 249 languages as of March 2026, following Google's 2024 expansion of 110 new languages. In late 2025, Google upgraded Translate with its Gemini language model, improving output quality on idioms, slang, and conversational content.

According to Intento's State of Translation Automation 2025, Google NMT appears among the top-performing solutions across nine of the eleven language pairs evaluated in both automated and human LQA assessment. For applications needing broad language coverage with solid quality across major pairs, Google's API is a strong default.

Pricing: $20 per million characters for the Basic (NMT) model. Advanced model pricing is higher and includes AutoML customisation features. A free tier is available.

Best for: Applications requiring broad language coverage, teams already in Google Cloud, general-purpose translation at scale.

Microsoft Translator Text API

Microsoft Translator is part of Azure Cognitive Services and integrates natively with the broader Microsoft ecosystem (Teams, Office, SharePoint). It supports over 130 languages and dialects.

In January 2025, Microsoft launched Microsoft Translator Pro for enterprise users, adding customised phrasebooks, expanded language coverage, and availability in US Government cloud. At $10 per million characters for standard usage, Microsoft offers the lowest per-character pricing among major NMT providers, along with the most generous free tier: 2 million characters per month with no expiration date.

Pricing: $10 per million characters for standard. Free tier: 2 million characters/month, no expiration.

Best for: Teams in the Microsoft ecosystem, applications requiring the lowest per-character cost, government and regulated-sector deployments with data residency needs.

Amazon Translate

Amazon Translate is an AWS-native translation service priced for volume. It integrates natively with S3, Lambda, and other AWS services, making it straightforward to embed in automated data pipelines. It supports over 75 languages.

Its Active Custom Translation feature allows organisations to supply parallel data to adapt the engine's output to specific terminology and domain. This reduces but does not eliminate the hallucination risk inherent in single-model systems.

Pricing: $15 per million characters for standard translation. Free tier: 2 million characters per month for the first 12 months.

Best for: High-volume pipelines in AWS infrastructure, applications where AWS ecosystem integration is more important than translation quality optimisation.

DeepL API

DeepL's API is the quality-focused option in the NMT category. It is particularly strong for European language pairs, where its training data produces more natural-sounding output than most competitors. DeepL now supports 33 languages, including Hebrew and Vietnamese added in 2024.

DeepL also launched DeepL next-gen in 2024, a purpose-built LLM specifically designed for translation that improves quality on longer texts and maintains enterprise security standards. Per Intento's State of Translation Automation 2025, DeepL next-gen appears among the top 14 solutions across language pairs evaluated.

Pricing: DeepL API Free: 500,000 characters/month at no cost, rate-limited, no document translation. DeepL API Pro: $25 per million characters, no monthly minimum, volume pricing for teams above 100 million characters/month. Consumer subscription plans (Starter, Advanced, Ultimate) are separate from API pricing.

Best for: Applications where European language quality is the primary requirement, legal and professional content requiring high fluency and naturalness.

LLM translation APIs: what changes with this category

LLM-based translation APIs work differently from NMT engines. Rather than a dedicated translation model, you are calling a general-purpose language model with a translation instruction. The trade-offs are real in both directions.

On quality: LLMs consistently outperform NMT engines on nuanced, domain-specific, and culturally sensitive content. They handle idioms, register, and brand voice better. Intento's 2025 evaluation found LLMs represent 89% of top performers across language pairs, with models like GPT-4.1, Claude Opus 4, and Gemini 2.5 Pro appearing as best-in-class across multiple pairs.

On cost and latency: LLMs are priced by token rather than character and are 10 to 100 times slower than NMT engines for the same content volume, per Intento 2025. For applications requiring sub-second response times or very high throughput, NMT remains the more practical choice.

The main LLM-based options for translation workflows:

OpenAI API (GPT-4.1): GPT-4.1 ranked first among single-agent solutions across 11 language pairs in Intento's 2025 human LQA evaluation. Pricing is $3.00 per million input tokens / $12.00 per million output tokens. Best for: high-quality professional translation where throughput is not the constraint.

Google Gemini API (Gemini 2.5 Pro/Flash): Gemini 2.5 Pro appears among the top performers across Arabic, German, Italian, Japanese, Korean, Dutch, Portuguese, Spanish, and Chinese in Intento's evaluation. Gemini 2.5 Flash offers lower latency at reduced cost. Best for: teams already in Google Cloud who want LLM-quality output.

Anthropic API (Claude Opus 4 / Sonnet 4): Claude Opus 4 ranks in the top tier for English to German, Japanese, Dutch, and Portuguese pairs. Claude Sonnet 4 is faster and cheaper. Best for: nuanced content where tone and register are primary.

DeepSeek V3 API: At approximately $0.27 per million input tokens, DeepSeek V3 is the most cost-efficient LLM API option. It appears as a top performer for English to Italian and English to Japanese in Intento's evaluation. Best for: high-volume LLM translation where cost is the primary concern.

How NMT and LLM APIs compare on quality, speed, and cost

Per Intento 2025, baseline NMT and LLM systems both fail professional standards when evaluated against specific requirements (terminology, tone, consistency, formatting). Customised systems (those with glossaries, terminology injection, or RAG) reduce errors by 80-90% relative to baseline. This applies to both categories.

API pricing comparison table (2026)

Pricing figures should be confirmed with each provider as they may vary by region, volume, and plan tier. LLM token prices are converted at approximately 2.83 characters per token following the methodology used in Intento's pricing analysis.

How to choose the right API for your use case

The most useful question to ask first is not "which API is best" but "what does my application actually need?" Three variables drive most decisions.

Latency requirements. If your application needs sub-second translation (real-time chat, live captioning, instant UI strings), NMT engines are the practical choice. LLMs are improving on latency but are still 10-100x slower at equivalent volume.

Content type and quality threshold. For general-purpose content at scale (product descriptions, support tickets, internal communications), Google, Microsoft, or Amazon will handle most use cases adequately at lower cost. For professional, client-facing, or regulated content where nuance and terminology accuracy are non-negotiable, LLM APIs or DeepL next-gen produce consistently stronger output.

Language pair. Google has the broadest coverage at 249 languages. DeepL is strongest for European pairs at 33 languages. LLMs have broad multilingual capability but performance varies by pair and is not always systematically benchmarked. For specific pair requirements, consulting Intento's State of Translation Automation 2025 language pair data is the most reliable starting point.

A practical heuristic for 2026: if you are building a new integration and quality is the primary concern, default to an LLM API and add glossary or prompt engineering customisation — both categories of model improve substantially with customisation. If you are optimising an existing NMT integration for cost or speed, stay on NMT and add glossary support if you have not already.

A note on multi-engine approaches

One pattern increasingly common in professional translation workflows is running multiple APIs and selecting the best output rather than routing all traffic through a single engine. Intento's 2025 evaluation found that multi-agent solutions tied with Google as the top performer by number of "best" rankings across language pairs, with the multi-agent approach achieving top-tier results in most languages.

The practical challenge with multi-engine approaches is integration complexity: managing multiple API keys, handling inconsistent output formats, and building the comparison logic. MachineTranslation.com's API handles this at the infrastructure level. It routes each translation through 22 models simultaneously, using its SMART consensus mechanism to identify the output the majority of models agree on, and returns that as the result alongside quality scores. Developers get the benefit of multi-model coverage without the overhead of managing each integration separately. The full model list includes Google, Amazon, DeepL, Microsoft, ChatGPT, Claude, Gemini, DeepSeek, Grok, and 13 others.

For teams who want to evaluate outputs across engines before committing to an integration decision, the MachineTranslation.com platform (no sign-up required) lets you run the same text through all 22 models in one interface.

FAQs

1. What is the best translation API in 2026?

There is no single best option. Google Cloud Translation has the broadest language coverage (249 languages) and consistent benchmark performance. DeepL produces the highest-quality output for European language pairs. Microsoft Translator offers the lowest per-character cost among NMT providers. For quality-critical content, LLM APIs (particularly GPT-4.1 and Claude Opus 4) now outperform NMT engines on most professional translation tasks according to Intento's State of Translation Automation 2025.

2. How do LLM translation APIs compare to NMT engines?

LLMs produce higher-quality output on nuanced, domain-specific, and professionally sensitive content. NMT engines are faster (sub-second vs. 10-100x slower for LLMs) and cheaper per character for standard usage. For real-time or very high-throughput applications, NMT remains more practical. For quality-critical work, LLM APIs have overtaken NMT on most benchmarks.

3. Is IBM Watson Language Translator still available?

No. IBM discontinued the Watson Language Translator service in June 2024. Teams using it need to migrate to an alternative. The most direct replacements are Google Cloud Translation or Microsoft Translator, depending on the language pairs and integration requirements.

4. What translation API has the lowest cost?

Microsoft Translator at $10 per million characters is the lowest-cost major NMT API in 2026. Among LLM APIs, DeepSeek V3 at approximately $0.27 per million input tokens is the most cost-efficient option, though pricing varies by deployment method and region.

5. Does DeepL have a translation API?

Yes. DeepL API Pro is priced at $25 per million characters with no monthly minimum. There is also a free API tier limited to 500,000 characters per month with rate limits and no document translation support. DeepL also offers DeepL next-gen, an LLM-based translation model available through their API for higher-quality output on longer texts.

6. What is the MachineTranslation.com API and how does it differ?

MachineTranslation.com's API routes each translation through 22 AI models simultaneously, including most of the APIs listed in this article, and returns the consensus output (the translation the majority of models agreed on) along with quality scores for each. Rather than choosing one engine, developers get a system that cross-checks its own output. This is useful when accuracy matters more than the lowest-cost-per-token. More details at MachineTranslation.com.