Description
The reasoning: -> effort: parameter (low | medium | high) does not appear to work, especially in custom agents. Regardless of value, results remain the same in terms of time, speed, and number of thinking tokens.

Tested Models

• gtp-oss-20
• gtp-oss-120
• grok-code-fast-1

Expected Behavior

• Clear differences between low, medium, and high reasoning.
• gtp-oss family normally shows large variation in thinking tokens:
  • low → ~58 tokens
  • medium → ~1500 tokens
  • high → 5000+ tokens

Actual Behavior

• No significant change between reasoning levels.
• Output corresponds roughly to medium level regardless of setting.
• OpenRouter activity logs show no increase in thinking tokens when switching from low to high (should be ~10x difference).

Steps to Reproduce

1. Run the same prompt with agents with different reasoning effort: low | medium | high.
2. Compare runtime, speed, and thinking tokens.
3. Observe logs for token usage.

Evidence

• Local runs of gtp-oss models confirm expected differences.
• Cloud service runs show no variation (all behave like medium).

Impact

• Impossible to control reasoning depth in custom agents.
• Misleads users expecting higher reasoning at higher effort levels.