Choosing an AI model

Translation workloads are usually high-volume and easy to evaluate on a fixed catalogue. Start with the inexpensive option, then compare terminology, placeholder preservation, JSON validity, latency, and total cost on your own messages before moving to a larger model.

Recommended defaults

OpenRouter: DeepSeek V4 Flash

Use deepseek/deepseek-v4-flash when cost is the priority:

ai_translation: {
  enabled: true,
  provider: 'openrouter',
  openrouter_api_key: process.env.OPENROUTER_API_KEY,
  model: 'deepseek/deepseek-v4-flash',
}

OpenRouter describes DeepSeek V4 Flash as an efficiency-optimized, high-throughput model with strong performance and a one-million-token context window. Its price depends on the routed provider and current discounts; use the live OpenRouter model page instead of copying a price into long-lived configuration.

This is also the default model shown in the browser translator.

OpenAI: GPT-5.6 Luna

Use gpt-5.6-luna when calling OpenAI directly:

ai_translation: {
  enabled: true,
  provider: 'openai',
  openai_api_key: process.env.OPENAI_API_KEY,
  model: 'gpt-5.6-luna',
}

The official OpenAI model page positions Luna for cost-sensitive, high-volume workloads. It is Translify's recommended OpenAI default. Use gpt-5.6-terra when an evaluation shows that the extra quality is worth the price, and gpt-5.6-sol for difficult, quality-first batches.

OpenRouter model slugs and OpenAI model IDs are not interchangeable. For example, use gpt-5.6-luna with the OpenAI provider and openai/gpt-5.6-luna when that model is available through OpenRouter.

Real Translify runs

These captures come from three real OpenRouter runs against the same 1,680-key project. They demonstrate what Translify records after each file and for the whole run; they are examples, not price guarantees. The model slug is not shown in the captured terminal output, so compare the billed amount with your own configured model rather than attributing these numbers to a specific model.

Spanish — 1,680 keys, 114,122 tokens, $0.0229

Danish — 1,680 keys, 97,008 tokens, $0.0224

Swedish — 1,680 keys, 52,125 tokens, $0.009429

Different target languages can produce very different completion-token counts. Treat cost per key as a measured range, keep batch_size stable during a model comparison, and review a representative sample before translating an entire repository.