The interesting part of Google’s latest-model API notes isn’t a new benchmark — it’s that temperature, top_p, and top_k are now deprecated and ignored on Gemini 3.6 Flash and 3.5 Flash-Lite, and future generations return an HTTP 400 if you send them at all.

If you run LLM ops in production, this quietly breaks a pile of assumptions. Half the eval harnesses I’ve seen treat temperature as a first-class sweep axis, and most orchestration wrappers pass a default temperature=0 on every call. Those wrappers start throwing 400s the moment a future model release lands — and nobody will have touched a line of their own code.

The deeper shift is philosophical. Google’s guidance is to get determinism from a system instruction with explicit rules, structured outputs to pin the format, and thinking_level to set reasoning effort — declarative constraints instead of a probability dial. Prefilled model turns are gone too; you move to the Interactions API. The provider is taking the low-level sampling surface away and handing you higher-level knobs it controls.

I think that’s the right call for agentic systems, where one stray high-temperature sample nine tool-calls deep is a reproducibility nightmare. But it moves determinism from something you own to something you rent. The “we froze temperature at 0 for the eval” story no longer means what it used to.

The HN thread is already full of people whose abstraction layers hardcode these params. If you maintain one, now’s the time to make temperature a no-op you strip, not a default you send. How many of your prod calls pass a sampling param you never actually tuned?