DeepL API vs Google Cloud Translation vs OpenAI: Real AI Translation Cost per 1 Million Characters in 2026

The short answer: dedicated machine-translation APIs and general-purpose LLMs bill in different units entirely, and converting between them changes the picture dramatically. On a straight per-character basis, DeepL API costs $25 per million characters, Google Cloud Translation's Basic NMT costs $20 (its newer LLM Translation mode also works out to about $20, and its custom-tuned Adaptive Translation to about $50), while an LLM like GPT-6 Sol, converted through a typical character-to-token ratio, costs roughly $3 per million characters — 6 to 8 times cheaper than the dedicated translation engines on this narrow metric. That gap is real but narrower in practice than it looks: at meaningful production scale, the dedicated NMT APIs' predictable per-character billing and purpose-built infrastructure can narrow the raw-price gap once prompt overhead, formatting preservation, and throughput are accounted for, though this article does not quantify how much of the gap those factors actually close.

What each vendor bills

DeepL API (REPORTED, cross-checked across multiple 2026 pricing guides, one figure conflicting). $25 per million characters, with a 500,000-character-per-month free tier. One source cites DeepL at $25/M consistent with the majority of sources reviewed; a separate consumer-facing DeepL Pro plan (not the API) starts at €5.99/month for 500,000 characters, a different product with a different billing structure that should not be confused with the developer API rate. DeepL is widely regarded as the strongest option for preserving document formatting (PDF, DOCX) during translation and for European-language quality specifically.

Google Cloud Translation (OFFICIAL, Google Cloud pricing documentation). Three distinct products under one pricing page, each priced differently: Basic NMT: $20 per million characters, with a 500,000-character-per-month free tier that never expires. LLM Translation mode (Gemini-powered): $10 per million input characters plus $10 per million output characters, no free tier, working out to roughly $20 per million characters translated end-to-end. Adaptive Translation (fine-tuned on your own data via a custom glossary): $25 input + $25 output per million characters, roughly $50 total — the most expensive of Google's three modes, priced for accuracy on domain-specific terminology rather than raw volume. A separate legacy AutoML custom-model tier scales down at high volume: $80/M for the first 500K–250M characters a month, falling to $60/M, $40/M, and $30/M at successively higher volume bands above 250 million, 2.5 billion, and 4 billion characters respectively — a different, older product from the current Adaptive Translation mode. Billing counts HTML tags as characters when present in submitted strings, a frequently-missed cost inflator for teams sending rich-text content without stripping markup first.

OpenAI (OFFICIAL for token rates, ILLUSTRATIVE for the character conversion). OpenAI has no dedicated translation-specific API or pricing tier; translation is done through general chat completion, billed per token like any other request. Using GPT-6 Sol at $2 input / $10 output per million tokens (established in this site's companion batch-pricing comparison) and an ILLUSTRATIVE conversion of 250 input tokens plus 250 output tokens per 1,000 source characters (an English-weighted estimate; the real ratio varies substantially by language and by which tokenizer a given model uses, and should be measured on your own content rather than assumed universal), translating 1 million characters costs roughly $3.00 — input contributes $0.50 and output $2.50 of that total, since output tokens are priced 5x input on this model.

Cost per volume

Formula: DeepL/Google Basic = characters (millions) × $20–$25; OpenAI = (characters ÷ 1,000 × 250 ÷ 1,000,000) × ($2 input + $10 output).

At every volume, OpenAI's token-converted cost undercuts both dedicated NMT APIs by roughly 7–8x on this narrow character-for-character basis, and Google's premium Adaptive Translation mode is the most expensive option in the table by a wide margin, reflecting its custom-glossary fine-tuning rather than raw throughput pricing.

Why the raw price gap narrows in practice

Independent industry guidance notes that LLM-based translation can be "sometimes up to 800 times less expensive" than NMT for short strings, and separately suggests that character-based translation APIs tend to become more predictable and cost-effective than token-priced LLMs at higher sustained volume. The rate cards themselves are linear with volume on every vendor in this article, so total monthly character count alone does not mechanically narrow the gap; what actually narrows it is a set of real but separately-varying cost factors this article's bare per-character math leaves out: per-call prompt overhead (system instructions, formatting guidance, and context resent on every LLM call, which a purpose-built NMT engine does not carry); request granularity and batching (many small calls carry proportionally more overhead than fewer large ones); retries for malformed or incomplete LLM output; formatting-preservation instructions, adding input tokens on every call for content that needs structure kept intact; output that runs longer than the source text, common in some language pairs and adding billed output tokens beyond the bare ratio; QA and reprocessing when LLM output needs a correction pass that NMT's more deterministic output needs less often; and throughput and latency requirements, where NMT APIs were reported to run 3 to 7 times faster than LLM-based translation, mattering independently of the per-character price itself. None of these factors has a single confirmed dollar value in the sources reviewed, and this article does not present any specific volume threshold as an automatic point where the economics flip; the 800x figure applies to a narrow best case (very short, simple strings where NMT's per-request overhead dominates), and the 6 to 8 times figure in this article's table is the bare rate-card comparison before any of the factors above are added back in for a specific real workload.

HTML and formatting: a real, easy-to-miss inflator

Google's Cloud Translation bills HTML tags as billable characters when present in submitted content. A string with heavy inline formatting (<b>, <span class="...">, nested divs) can have its markup account for a meaningful share of total billed characters, inflating the real per-translated-word cost well above what the plain-text character count would suggest. Stripping markup before submission and re-applying it after translation is the standard mitigation, but it adds engineering overhead unmodeled in any of the headline rates above.

Sensitivity

1. Which OpenAI model tier is used. A frontier model at 2–3x GPT-6 Sol's rate would still likely undercut DeepL and Google Basic on raw per-character cost, but by a narrower margin; a cheaper tier model would widen OpenAI's advantage further.
2. Character-to-token ratio. Character-to-token ratios vary materially by language, script, content and tokenizer; measure them on representative source and target text rather than assuming the 250-per-1,000 English-weighted ratio used here applies universally.
3. HTML/markup density. Content with heavy inline formatting inflates Google's billed character count without inflating the actual translated word count; this article's sources establish this specifically for Google Cloud Translation and do not establish the same mechanism for DeepL's API.
4. Which operational factors apply to your pipeline. Per-call prompt overhead, request granularity, retries, formatting-preservation instructions, longer-than-source output, QA/reprocessing, and throughput needs each narrow the raw LLM-versus-NMT price gap by an amount specific to your own workload; this article does not tie any of them to a single volume threshold.

Budgeting traps

• Comparing OpenAI's bare token-converted rate to DeepL's or Google's rate without adding prompt overhead. The bare $3/M-character figure understates real LLM translation cost once system instructions and formatting guidance are included in every call.
• Ignoring HTML tags as billable characters on Google Cloud Translation. Strip markup first, or budget for the inflated character count.
• Assuming DeepL Pro's consumer per-month pricing applies to API usage. The developer API bills separately, at a different rate structure, from the consumer subscription product.
• Choosing Google's Adaptive Translation for raw-volume throughput. It is priced for custom-glossary accuracy, at roughly 2.5x Google's own Basic NMT rate, not for being the cheap high-volume option.

What to ask before you buy

Measure your actual character-to-token ratio on a representative sample of your own content, in your target languages, before trusting the 250-per-1,000 English-weighted estimate used here. Ask whether your pipeline sends markup-heavy or plain-text content to Google Cloud Translation specifically, since HTML tags billing as characters can meaningfully change the real per-word cost.