Choosing an AI Provider
Batch Translate with AI works with any OpenAI-compatible API. Nine providers are built in — you just pick one, get an API key, and paste it in. No other configuration needed.
Quick recommendation
| If you want… | Use |
|---|---|
| Cheapest option for bulk translation | DeepSeek |
| Best overall quality | OpenAI (gpt-4o or gpt-4o-mini) |
| Cheapest multi-language runs | OpenAI or DeepSeek (prompt caching) |
| Fastest responses | Groq |
| No account, many models | OpenRouter |
| European data residency | Mistral AI |
Multi-language efficiency (prompt caching)
When you translate one document into several languages, the source text has to reach the model for each language. Providers with prompt caching reuse the source from cache for languages 2…N at a large discount instead of charging full price every time.
In the app, caching-capable providers are labelled “Cost-efficient for multi-language.” The translation engine still works on any provider — it just won’t get the discount on ones without caching, and the settings dialog tells you which is which.
| Caching support | Providers |
|---|---|
| Yes (cost-efficient for multi-language) | OpenAI, DeepSeek, OpenRouter* |
| No (still works, no multi-language discount) | Groq, Mistral, Together AI, xAI, Fireworks, Cerebras |
* OpenRouter caching depends on the underlying model you route to.
All supported providers
| Provider | Strengths | Free tier | Sign up |
|---|---|---|---|
| OpenAI | Best quality, reliable, prompt caching | No (credits on signup) | platform.openai.com |
| DeepSeek | Extremely low cost, strong quality, context caching | Yes (limited) | platform.deepseek.com |
| OpenRouter | 100+ models from one key, caching varies by model | No | openrouter.ai |
| Groq | Very fast inference, good for high volume | Yes (rate limited) | console.groq.com |
| Mistral AI | Strong multilingual, EU-based | Yes (limited) | console.mistral.ai |
| Together AI | Wide model selection, competitive pricing | Yes ($1 credit) | api.together.ai |
| xAI (Grok) | Grok models, good reasoning | No | console.x.ai |
| Fireworks AI | Fast, cost-effective | Yes ($1 credit) | fireworks.ai |
| Cerebras | Ultra-fast inference | Yes (limited) | cloud.cerebras.ai |
Cost considerations
All providers charge per token — roughly 1 token per 0.75 words. Translation output is usually a little longer than the source, and some languages (German, Finnish) use more tokens per character than English.
- A 10-page document into one language uses very roughly 8,000–20,000 tokens total
- A 300-page book into one language can run into the low millions of tokens
- At DeepSeek’s rates a large book typically costs a few dollars; premium models cost more
- Translating into N languages with a caching-capable provider costs far less than N× the single-language cost
Use Estimate size before large jobs, and watch the live projection during a run. The app deliberately does not display prices — multiply the token figure by your provider’s current published rate.
Using a custom endpoint
If you use a self-hosted model (Ollama, LM Studio, LocalAI) or a provider not in the list, select Custom (OpenAI-Compatible) from the provider dropdown and enter your endpoint’s base URL manually. The endpoint must be OpenAI-compatible (/chat/completions). See Finding Your Base URL.
Switching providers
You can switch providers at any time from the key icon. Base URL, model, and the caching flag update automatically from the preset; your API key is stored securely on device.