Translation Task Cost: GPT-4o vs DeepL vs Google Translate

The 2024 Inversion

Five years ago, dedicated translation APIs (DeepL, Google Translate) were significantly cheaper per token than asking GPT-4 to translate. With GPT-4o pricing, that has flipped.

Per-1M-tokens cost

Service	Approx $ per 1M tokens (translation)
GPT-4o (input + output)	$12.50
GPT-4o mini	$0.75
Claude Haiku 4.5	$6.00
Gemini 2.5 Flash	$2.80
DeepL Pro	~$100
Google Translate API	~$80

GPT-4o mini at $0.75 per 1M tokens is 130x cheaper than DeepL Pro for the same workload.

The quality caveat

DeepL still produces more natural output for single-document, no-context translation in well-resourced language pairs (English ↔ German, French, Japanese). Where LLMs win:

• Glossary / brand-term preservation — a system prompt instructing "always translate 'DevToolbox' as 'DevToolbox' (do not localize)" works first time. DeepL requires per-account glossaries.
• Domain-specific terminology — legal, medical, technical translations benefit from a few-shot example in the prompt.
• Tone preservation — "translate maintaining the casual, second-person tone" is a one-line instruction.
• Long-form context — translating chapter-by-chapter where earlier translations inform later choices.

Practical math

To translate a 100-page document (~50,000 words ≈ 67,000 tokens English → ~80,000 tokens Japanese):

• GPT-4o mini: 67K input × $0.15/1M + 80K output × $0.60/1M = $0.058
• GPT-4o: 67K input × $2.50/1M + 80K output × $10/1M = $0.97
• Claude Haiku 4.5: 67K × $1/1M + 80K × $5/1M = $0.47
• DeepL Pro: 80K × $100/1M ≈ $8.00

GPT-4o mini for 6 cents vs. DeepL for $8 is a 130x savings.

When to still use DeepL

• Compliance: certified translations for legal filings.
• Latency: DeepL response time is typically faster than LLM completion.
• Predictability: DeepL has zero hallucination risk for pure translation.