The cheapest translation runs I measured cost 0.08¢ per 1,000 words: Qwen3.8 Flash and GPT-6 Luna, at $0.82 and $0.83 per 1,000 tasks. The median across all 326 models is 0.79¢, so the cheapest options cost roughly a tenth of a typical model. Price per token is only part of the bill, because tokenizer differences, output length and hidden reasoning tokens can move the final cost well away from the list price.
Cost per 1,000-word translation
My translation task uses 1,017 words of The Prince, translated into Spanish. I report cost per task and cost per 1,000 tasks. The cheapest models in the set:
- Qwen3.8 Flash: 0.08¢ per task ($0.82 per 1,000 tasks). Measured.
- GPT-6 Luna: 0.08¢ per task ($0.83 per 1,000 tasks). Measured.
- Llama 4 Maverick: 0.11¢ per task ($1.10 per 1,000 tasks). Estimated.
- Claude Haiku 5.5: 0.13¢ per task ($1.26 per 1,000 tasks). Measured.
- GPT-5.4 Nano: 0.20¢ per task ($2.01 per 1,000 tasks). Estimated.
The upper end looks very different:
- Claude Sonnet 5.5: 2.6¢ per task ($25.52 per 1,000 tasks). Measured.
- Claude Opus 5.5: 5.1¢ per task ($51.05 per 1,000 tasks). Estimated.
- Claude Fable 5.1: 13¢ per task ($127.62 per 1,000 tasks). Estimated.
- Kimi K3: 18¢ per task ($182.66 per 1,000 tasks). Estimated.
The table shows list prices next to token counts for the cheapest models. Prices are USD per 1M tokens as of 2026-10-11.
| Model | Input / output price (per 1M) | Cost per 1,000 words | Input / output / reasoning tokens | Basis |
|---|---|---|---|---|
| Qwen3.8 Flash | $0.15 / $0.47 | 0.08¢ | 1,386 / 1,166 / 136 | Measured |
| GPT-6 Luna | $0.10 / $0.50 | 0.08¢ | 1,330 / 1,328 / 67 | Measured |
| Llama 4 Maverick | $0.19 / $0.65 | 0.11¢ | 1,375 / 1,156 / 136 | Estimated |
| Claude Haiku 5.5 | $0.10 / $0.50 | 0.13¢ | 1,832 / 2,153 / 0 | Measured |
| GPT-5.4 Nano | $0.20 / $1.25 | 0.20¢ | 1,330 / 1,328 / 67 | Estimated |
See the translate 1,000 words task page for the full ranking.
Budgets per 1,000 tasks
One thousand tasks of 1,017 words each is roughly one million words. Here is what that costs, using the per-1,000-task figures from my data:
| Model | Cost per 1,000 tasks | Basis |
|---|---|---|
| Qwen3.8 Flash | $0.82 | Measured |
| GPT-6 Luna | $0.83 | Measured |
| Llama 4 Maverick | $1.10 | Estimated |
| Claude Haiku 5.5 | $1.26 | Measured |
| GPT-5.4 Nano | $2.01 | Estimated |
| Claude Sonnet 5.5 | $25.52 | Measured |
| Claude Opus 5.5 | $51.05 | Estimated |
| Claude Fable 5.1 | $127.62 | Estimated |
| Kimi K3 | $182.66 | Estimated |
The gap between the cheapest and the most expensive model in this list is large, but the absolute amounts stay small for most projects. Use the calculator to plug in your own word counts.
Why tokens and reasoning change the bill
List price per token is the starting point. Three things move the final number:
- Tokenizer differences. For the same English document, Anthropic's tokenizer produces 1.39 times as many tokens as OpenAI's (OpenAI = 1.00). Claude Haiku 5.5 used 1,832 input tokens per translation against 1,330 for GPT-6 Luna, though both list input at $0.10 per 1M tokens.
- Output length. Claude Haiku 5.5 produced 2,153 output tokens against 1,328 for GPT-6 Luna. Output tokens are priced higher than input tokens for every model in my data.
- Reasoning tokens. These are billed as output tokens. GLM 5.3 Flash used 8,761 reasoning tokens and still cost 0.53¢ because its per-token prices are low. Mistral Large 4 used 10,816 reasoning tokens and cost 2.6¢.
Reasoning mode varies by model. Claude Sonnet 5.5 and Gemini 3.8 Flash reason always, some models reason by default, and Llama 4 Maverick and Qwen3.8 Flash are listed with no reasoning mode. Their translation runs still show 136 reasoning tokens each. If you need predictable cost, check reasoning tokens in your own runs before you commit a budget.
Proofreading follows a similar pattern
For one 500-word page, GPT-6 Luna costs 0.04¢, the lowest in the set, and Kimi K3 costs 6.2¢. The median across all 326 models is 0.26¢. The ranking is not identical to translation: GLM 5.3 Flash is second for proofreading at 0.05¢, while Qwen3.8 Flash drops to fourth at 0.10¢. See the proofread text task page for details.
Quality caveats
Cost is half the decision.
- I did not score translation quality. The figures measure tokens and price, not how accurate or natural the Spanish is. Test outputs on your own text before switching providers.
- LMArena is not a translation benchmark. It measures general chat performance. GPT-6 Luna scores 1443, GPT-5.4 Nano scores 1401, and Llama 4 Maverick scores 1327, the lowest in my data. Claude Haiku 5.5 and Qwen3.8 Flash have no listed score. None of these tells you how a model handles literary Spanish.
- Consistency across long projects. My budgets assume separate 1,000-word tasks, but a real project needs consistent terminology across chapters. Test that over a full project, not one page.
- Measured versus estimated. Measured rows are single runs with default settings via OpenRouter, with output capped at 16,000 tokens. Estimated rows are approximate. Of the five cheapest, Llama 4 Maverick and GPT-5.4 Nano are estimated.
For a second check on quality, budget for human review of a sample. See plans for plan options and the methodology page for the full method.
FAQ
What is the cheapest AI for translation per 1,000 words?
In my measurements, Qwen3.8 Flash and GPT-6 Luna both cost 0.08¢ per 1,000-word translation, or $0.82 and $0.83 per 1,000 tasks. The median across all 326 models is 0.79¢.
Does a lower price per token always mean a lower translation cost?
No. Tokenizers differ and reasoning tokens are billed as output. Claude Haiku 5.5 has the same $0.10 input price as GPT-6 Luna but used more input and output tokens, and cost 0.13¢ against 0.08¢.
Are these cheap models good enough for professional translation?
I did not measure translation quality, so I cannot say. Test any model on your own documents and have a human review a sample before using it for client or publication work.