Google brings Gemini 3.6 Flash and Gemini 3.5 Flash-Lite to general availability

Google has moved Gemini 3.6 Flash and Gemini 3.5 Flash-Lite out of preview and into general availability on the Gemini API as of July 21, 2026, while simultaneously deprecating the classic sampling parameters that have shaped developer-side tuning since the API's early days.

What's new

The Gemini API changelog describes the release as delivering "stable, production-ready versions of our latest 3.x Flash models." Two models are affected:

• Gemini 3.6 Flash (gemini-3.6-flash): according to Google, it "features improved token efficiency and code/agentic planning capabilities at a lower price point than 3.5 Flash, resolving developer feedback around output verbosity." That framing points directly at a common complaint about Flash-tier models — that they run up token counts on verbose, meandering output — and positions 3.6 Flash as the more disciplined, agent-friendly successor.
• Gemini 3.5 Flash-Lite (gemini-3.5-flash-lite): Google calls it a "low-latency, highly cost-effective subagent option designed for high-volume automation," aimed squarely at pipelines that fan out many small, cheap model calls rather than a handful of expensive ones.

Alongside the GA push, Google deprecated the temperature, top_p, and top_k sampling parameters across the Gemini API. Any integration still setting these directly on requests needs to migrate before they stop having effect.

Context

Flash-tier models exist to handle the high-volume, latency-sensitive side of the workload that flagship models like Gemini 3.1 Pro aren't built for economically. Google has been iterating on this tier quickly, and general availability — rather than preview status — is what lets enterprise customers actually build production SLAs on top of a model rather than treating it as an experiment. The parameter deprecation lands in the same release, which is notable: Google is bundling a routine capability upgrade with an API-wide behavior change, forcing anyone touching the Flash family to also handle the sampling-parameter migration at the same time.

Why it matters

On the surface this is a routine GA graduation, but two details make it more consequential than a typical minor-version bump. First, positioning 3.6 Flash explicitly around "agentic planning capabilities" signals that Google is tuning even its cheapest tier for tool-calling and multi-step agent workloads, not just chat completions — consistent with the industry-wide shift toward agents as the default consumption pattern for API traffic. Second, deprecating temperature, top_p, and top_k removes long-standing manual controls in favor of whatever internal tuning Google's models now default to, mirroring a similar move by Anthropic (which returns a 400 error for non-default sampling parameters on newer Claude models). That convergence suggests frontier labs increasingly see manual sampling control as a legacy interface rather than a feature worth preserving, since their newer models are tuned to perform best under fixed, internally-managed sampling behavior.