Google has moved two models to general availability: gemini-3.6-flash and gemini-3.5-flash-lite. Both are now stable endpoints rather than previews.

What changed

Gemini 3.6 Flash ships with improved token efficiency and better code and agentic planning at a lower price point than the model it replaces. The token-efficiency claim matters more than it sounds: on long agentic runs the bill is driven by how many tokens the model spends thinking, not by the sticker price per million.

Gemini 3.5 Flash-Lite is positioned as a low-latency, cost-effective subagent option — the model you call thousands of times inside a larger system rather than the one a user talks to.

The breaking change

Shipping alongside GA is a change that will break existing code: the temperature, top_p and top_k sampling parameters are deprecated for both models. Code that passes them needs review. This is a real shift in posture — those three knobs have been standard across generative APIs since the beginning, and removing them means the model decides its own sampling behaviour.

Why it matters

Deprecating sampling controls trades user tuning for vendor control. For most applications the defaults were already fine, and fewer knobs means fewer ways to misconfigure a deployment. For teams that spent real effort tuning temperature per use case, that work is now obsolete.

Notably absent from this release is Gemini 3.5 Pro, which has now slipped past several expected windows. Google shipped the fast, cheap tier and left the flagship for later.

What to do about it

Audit any code path that sets temperature, top_p or top_k against these two models and plan to remove it. Deprecation is not immediate removal, but it is a commitment, and silently ignored parameters are worse than errors — output changes and nothing tells you why.

Where determinism actually matters, the replacement is structural rather than parametric: constrain the output format, validate what comes back, and retry on failure. That approach survives model upgrades in a way that a tuned temperature value never did.

The pattern across vendors

Google is not alone here. The industry has been steadily removing low-level controls in favour of models that manage their own inference behaviour, and reasoning models made most sampling tweaks meaningless anyway. Expect the surface area of these APIs to keep shrinking.