Gemini 4 Hydrogen ignores temperature, mandates thinking_level

Decoupling reasoning models from classic sampling knobs is becoming standard practice as inference engines replace stochastic token selection with discrete compute budgets. While discrete thinking tiers simplify agent harnesses, they eliminate granular developer control over per-token latency and generation costs.

• –Requests providing temperature, top_p, or top_k silently drop those parameters, while setting thinking_budget or MINIMAL returns an HTTP 400 error.
• –Reasoning depth scales across four discrete tiers: LOW for classification and routing, MEDIUM for standard coding, HIGH for agent tool loops, and XHIGH for multi-file software engineering.
• –Unlike Carbon and Argon which baseline at MEDIUM, Hydrogen defaults to HIGH, enforcing reasoning depth out of the box.
• –Agent harnesses that rewrite or truncate prompt histories must explicitly set prefix_mismatch_behavior to DROP to prevent reasoning state desyncs.