Language support question: Prompt API seems to understand non-en/es/ja inputs when inputLanguage is omitted (e.g., hu)

[scsinfo]

Hi team - I have a language-support question from a production integration.

Docs mention limited/explicit support for inputLanguage (e.g. en/es/ja). However, in practice we observe that if we omit inputLanguage, and provide sharedContext / context / prompt text in another language (e.g. Hungarian hu), the model often appears to understand and respond almost correctly.

Real-world context (why this matters)

We ship a WordPress plugin used by agencies. Editors often work in non-English (e.g. hu), while we keep model output in English for consistency and translate separately. In production this means prompts/context can be multilingual, and we need to know what language behavior is guaranteed vs. accidental.

What we observe

• inputLanguage not provided
• sharedContext and/or prompt content is hu (or other languages outside en/es/ja)
• The model produces Hungarian output based on the non-English input

Questions

1. Is this behavior supported/expected, or is it just “best effort” and not guaranteed?
2. When inputLanguage is omitted, is there any automatic language detection happening, or is this purely model-dependent behavior?
3. If we rely on this, what are the failure modes you've seen (e.g., worse quality, higher latency, silent misinterpretation)?
4. Would you recommend always setting inputLanguage to a "closest supported" value, or leaving it unset when inputs are multilingual?

Happy to provide anonymized minimal examples if useful.

p.s. Translator + Language Detector are extremely fast - great job.