Translation sits in an awkward spot in late 2026. The task is one of the oldest commercial uses of AI, but the market has split into two columns that barely share a currency. Dedicated machine translation engines bill per character. General-purpose LLMs bill per token. List prices on both sides have moved in opposite directions over the past year, and the gap between them is now large enough to change buying decisions.
This piece prices out both columns using published list prices, then looks at what the per-unit math hides.
The dedicated MT menu
All four major dedicated providers bill on translated characters and publish flat per-million rates. As of November 2026:
- Azure Translator charges roughly $10 per million characters for standard text translation, with a free tier of 2 million characters per month. Third-party trackers like better-i18n and Activepieces cite the $10 rate; Microsoft's own pricing page now renders per-unit rates dynamically and confirms the 2 million character free tier.
- Amazon Translate charges $15 per million characters for standard text and batch document translation across 75 languages, with 2 million characters per month free for the first 12 months. Real-time DOCX translation costs double at $30 per million, and Active Custom Translation, which reranks output toward your parallel-data examples, runs $60 per million.
- Google Cloud Translation charges $20 per million characters for default NMT after a monthly 500,000 character credit. Custom AutoML models are volume-tiered from $80 down to $30 per million. Document translation is priced per page at $0.08. Notably, Google now sells LLM translation inside the same product: Translation LLM bills $10 per million input characters plus $10 per million output characters, roughly the same effective price as NMT, while Adaptive Translation, the fine-tune-on-your-data variant, runs $25 plus $25.
- DeepL API is the premium corner of this column. Comparison articles consistently place it at $25 per million characters on the Pro plan plus a monthly base fee, around $32.50 per million on the Growth plan, with fewer supported languages than Google's 130+ but a reputation for the most natural output on European language pairs. Its enterprise features (glossaries, style rule lists, translation memory) are tiered by plan, per DeepL's own pricing page.

So the dedicated column clusters between $10 and $32.50 per million characters, with custom and adaptive variants reaching $50 to $60.
The token column, converted
LLM APIs do not publish per-character prices, so they need converting. At roughly four characters per token for English text, translating a million characters (assuming output of similar length) costs:
- GPT-6 Luna, at $0.10 per million input tokens and $0.50 per million output tokens, lands near $0.15 per million characters. CloudInsight's comparison derives the same arithmetic from OpenAI's official list prices, and The Daily Brief independently repeats the model rates.
- Claude Haiku 5.5 sits in the same band at $0.10 to $0.60 input (tiered by prompt size) and $0.50 output per million tokens.
- Gemini 3.8 Flash costs $0.75 input and $3.75 output per million tokens, about $1.13 per million characters, with a documented price increase to $1.50/$7.50 scheduled for January 2027.
- GPT-6 Sol and Claude Sonnet 5.5, both at $2/$10 per million tokens, work out to roughly $3 per million characters, still below every dedicated engine.
Both OpenAI and Anthropic offer batch APIs at roughly half the standard rate, which pushes the cheap tier under $0.10 per million characters.
On raw list price, a small LLM now translates text at one to two orders of magnitude below the per-unit cost of Azure, Amazon, Google, or DeepL. That is the headline number, and like most headline numbers in this market, it is only true after you answer what you are actually buying.
What the unit price hides
Three costs do not appear in either price list.
Token density varies by language pair. The four-characters-per-token conversion is an English approximation. CloudInsight's comparison warns that Chinese and English carry different token densities per character and recommends running a small batch to derive your own factor before trusting any cross-provider math. Per-character MT pricing has no equivalent landmine; the meter is the meter.
Terminology control is a product feature in one column and a project in the other. Dedicated engines sell determinism directly: glossaries and custom models at Google, parallel-data reranking at AWS, glossaries, style rules and translation memory at DeepL. With an LLM you assemble prompt-level glossaries yourself, and DeepL's own quality blog post argues the point sharply: general-purpose LLMs "trade away determinism, compliance, and consistency" because their output is probabilistic. That is a vendor talking, but the architectural claim is fair.
Quality measurement is the tie-breaker, and it is close. DeepL says its March 2026 benchmarking, 48,000 blind tests with language experts across 16 language pairs, showed it winning 94% of test groups against five competitors including the largest LLMs. The same post concedes that frontier LLMs at high-reasoning settings can be competitive on specific pairs. Academic work points the other way on context: an arXiv study comparing DeepL and Supertext on unsegmented documents found differences that segment-level tests miss, larger at document level, and cites the Intento report finding DeepL and GPT-class models consistently at the top among engines. A 2025 evaluation of technical content measured GPT-4o at 49.32 BLEU, and no single tool led across all pairs. High-reasoning settings also add latency, which is exactly what the cheap token tier does not buy you.

Reading the two columns together
The market is converging on both sides. Google sells LLM translation at its NMT price inside a per-character product, and DeepL has been shipping a specialized translation LLM behind per-character billing since 2024, reporting 1.4x to 1.7x blind-test quality gains over its prior engine on key pairs. The durable difference is the unit of accountability. Per-character contracts are auditable by anyone; per-token costs depend on your prompt, your segmentation, and your language pair.
The practical split looks like this:
- If you translate high-volume, steady-state content with fixed terminology (support docs, catalogs, UI strings), a dedicated engine at $10 to $25 per million characters remains the defensible default, because glossary enforcement and stable output are included in the price.
- If you translate uneven, context-heavy material (marketing, long documents, low-resource pairs) and already run an eval pipeline, small LLMs at $0.15 to $3 per million characters undercut the dedicated column outright, and batch pricing widens the gap.
- The hybrid is already normal: LLM for drafts and hard segments, dedicated MT with a locked glossary for volume. Google's Adaptive tier at $50 effective per million characters exists precisely because some buyers want LLM quality with an MT-shaped bill.
What has changed in the past year is the floor. Once a Luna-class model crosses the quality bar for your language pair, the per-character incumbents are no longer competing on price; they are competing on governance features, and they are charging a multiple for them. Whether that multiple is worth it depends entirely on whether your volume can tolerate an occasional creatively translated product name.
Sources
- Google Cloud Translation pricing: https://cloud.google.com/translate/pricing
- Amazon Translate pricing: https://aws.amazon.com/translate/pricing/
- Azure Translator pricing: https://azure.microsoft.com/en-us/pricing/details/translator/
- DeepL pricing: https://www.deepl.com/en/pricing
- DeepL, "Should you still care about translation quality?": https://www.deepl.com/en/blog/should-you-still-care-about-translation-quality
- DeepL press release, next generation LLM: https://www.deepl.com/en/press-release/deepl-launches-next-generation-llm-that-outperforms-gpt-4-google-and-microsoft-for-translation-quality
- CloudInsight, AI Translation API Comparison 2026: https://cloudinsight.cc/en/blog/ai-translation-api
- The Daily Brief, Machine Translation Costs $20 on Google, Under $1 With Small LLMs: https://www.beri.net/article/machine-translation-pricing-per-character-vs-per-token-llm-nmt-post-editing-cost
- better-i18n, Best AI Translation Tools in 2026: https://better-i18n.com/en/blog/best-ai-translation-tools-2026
- SimpleLocalize, AI translation cost comparison: https://simplelocalize.io/blog/posts/ai-machine-translation-cost-comparison
- Activepieces, Google Translate API Pricing: https://www.activepieces.com/blog/google-translate-api-pricing-cost-per-million-2026
- BuildMVPfast, Translation API Pricing Comparison: https://www.buildmvpfast.com/api-costs/translation
- arXiv, A comparison of translation performance between DeepL and Supertext: https://arxiv.org/html/2502.02577v2
- OpenAI pricing: https://openai.com/api/pricing/
