ClaudeCode in Slovenian, I wanted to see if I was sacrificing performance for convenience. After analyzing my own session transcripts and cross-referencing them with recent LLM research, the gap is much smaller than the community lets on.
Performance Analysis #
The internal logic of large models appears to be largely language-neutral. Research indicates that models process meaning in a shared conceptual space that leans toward English, meaning non-English input essentially acts as a thin translation layer.
Frontier Models: The performance drop is negligible. While older models struggled, current benchmarks show major languages retaining 97-98% of English performance. Even in less-represented languages, the difference in pragmatic evaluation is often only a few percentage points.Small Models: This is where the risk lies. Smaller, distilled models show a massive drop-off in non-English proficiency (sometimes losing nearly 50% of their effectiveness). If you're using a "Haiku-class" model, English is still a necessity.Code-Switching: Mixing English technical terms into a non-English prompt actually improves comprehension in some cases, likely because the model can anchor the request to high-density English training data.
The Token Tax #
The real cost of avoiding English isn't quality—it's efficiency. Non-English text is significantly more "expensive" in terms of token consumption.
Token Expansion: In my tests, Slovenian text consumes roughly 1.88x more tokens than the equivalent English translation.Context Window Impact: While not a dealbreaker for short prompts, this expansion eats into your context window faster and increases costs if you are paying per token.
Practical Takeaways for AI Workflows #
If you are building a professional AI workflow or using an LLM agent for coding, here is the breakdown:
-
For Complex Logic: Use the biggest model available (Claude Opus or GPT-4o). At this scale, you can prompt in your native language without a noticeable hit to reasoning quality.
-
For Budget/Speed: If you are deploying a small model for a specific task, stick to English to avoid the "intelligence drop" seen in smaller parameter counts.
-
For Technical Tasks: Don't bother translating technical terms. Keep "deployment," "callback," or "middleware" in English even if the rest of your prompt is in another language; it actually helps the model.
Ultimately, the friction of translating your thoughts into English often outweighs the marginal gain in model accuracy. Unless you're hitting token limits or using tiny models, prompt in the language that lets you express your logic most clearly.
Next How to Implement Human-in-the-Loop for AI Emails →