qwen/qwen3.8-max-0902: which native reasoning_effort do "minimal" and "high" map to?

[mauricekleine]

Hey! I run Nonobench, a benchmark that has LLMs solve nonogram puzzles through OpenRouter. A reader spotted something odd in our Qwen3.8 Max results, and I can't answer it from the API alone.

What I see

• /api/v1/models lists supported_efforts: ["xhigh", "high", "medium", "low", "minimal"] for qwen/qwen3.8-max-0902.
• Qwen's own docs list three levels for Qwen3.8-Max: low, medium and xhigh (default).
• Your reasoning docs say unsupported levels map "to the nearest supported level". But the model lists all five as supported, and the Alibaba endpoint has reasoning: null, so I can't tell which levels are native and which are mapped.

What our runs suggest

Same 30 puzzles, one attempt each, all served by Alibaba. Average tokens per puzzle:

Sent	5x5	10x10	15x15	Solved
minimal	3.1k	36.7k	39.9k	18/30
low	3.1k	34.4k	45.1k	18/30
medium	4.2k	38.5k	38.3k	14/30
high	4.7k	30.1k	55.9k	18/30
xhigh	5.6k	32.6k	58.6k	18/30

That looks like three settings, not five: minimal → low, and high → xhigh. high sits between medium and xhigh, so "nearest" could go either way.

Questions

1. Which native effort do minimal and high map to for this model?
2. Could supported_efforts list only native levels, or flag the mapped ones? Then anyone running effort ladders through OpenRouter knows which levels are real before paying for them.
3. Is there a way to see the native effort you forwarded, e.g. in the generation metadata?

Happy to share generation IDs if that helps. Thanks for looking!