Gemini kills temperature controls, and builders are annoyed
Google quietly deprecated temperature, top_p, and top_k in the latest Gemini models. These parameters are now silently ignored, and future model generations will return an HTTP 400 error if you supply them. Claude Sonnet 5 also dropped temperature. Google's suggested replacement is 'define a system instruction with explicit rules,' which the HN thread immediately called out as not a real substitute for stochastic sampling control.
This is a pattern of providers pulling back developer control in favor of internal optimization. The framing from Google is about 'determinism,' but the practical effect is that if you've built workflows that depend on low-temperature outputs for reliability, you need to rewrite your prompting strategy. The 'just use a system prompt' advice received open skepticism.
The deeper tension: as models get more capable, providers seem to be standardizing the interface downward, removing knobs that advanced users depend on. This is fine for casual users and bad for builders who've tuned production systems around specific parameter ranges.
So what?
Audit any production pipeline that relies on temperature or top_p settings now. If you're on Gemini or Sonnet 5, those calls will break. More broadly, watch for this trend across providers: the more control you need, the more you should prefer open-weight models where you own the inference stack.