Key Takeaways
- Gemini 3.6 Flash and Gemini 3.5 Flash-Lite are now generally available and positioned for production use.
- For these models, Google says sampling parameters such as temperature, top_p, and top_k are deprecated and ignored.
- The migration guidance emphasizes removing prefilled model turns and updating code to the Interactions API.
What happened
Google’s Gemini API documentation has been updated with details on two new generally available models: Gemini 3.6 Flash and Gemini 3.5 Flash-Lite. The company says both are ready for production use and support a 1M token context window, 64k maximum output tokens, thinking, and built-in tools including Computer Use.
The documentation frames the release as both a model update and an API transition. It includes quickstart examples for Python, JavaScript, and REST, and it points developers to migration steps for each model.
The most consequential API change in the source is a policy shift around sampling and turn structure. Google says temperature, top_p, and top_k are deprecated and ignored on these models. The company also says future Gemini model generations will return an HTTP 400 error if those parameters are supplied.
The same documentation says prefilling model turns is no longer supported. If the last non-empty turn in a request is a model turn, the API returns a 400 error. For developers who previously used that technique to suppress preambles or force a format, Google recommends using system instructions or structured outputs instead.
The release notes also describe what changes in each model. Gemini 3.6 Flash is presented as the stronger general-purpose option, with gains in agentic and multimodal work, fewer reasoning steps, fewer tool calls, improved code generation, and better instruction following. Gemini 3.5 Flash-Lite is described as the fastest and lowest-cost model in the 3.5 family, with a focus on high-throughput execution and document parsing.
The documentation says Gemini 3.6 Flash also becomes the default model powering the Antigravity agent in Gemini Managed Agents, and that this default can be changed through a new API field.
Why it matters
For developers building on Gemini, the story is less about a single model launch and more about how Google wants those models to be used. Removing temperature, top_p, and top_k changes a familiar part of LLM tuning. Teams that relied on those controls for output variability, determinism, or prompt experimentation will need to adjust their workflows.
Google’s guidance suggests that determinism should now come from system instructions that define explicit rules for the use case, rather than from sampling parameters. That is an important shift for application design because it moves more of the control surface from probabilistic knobs toward instruction design and structured output handling.
The turn-validation change matters too. Applications that used prefilled model responses as a workaround for formatting, preambles, or partial outputs will need to move to supported patterns. Google specifically points developers to system_instruction and structured outputs in the Interactions API.
The update also signals where Google sees the newer models fitting. Gemini 3.6 Flash is positioned as the better choice for complex agentic and multimodal tasks, while Gemini 3.5 Flash-Lite is aimed at throughput-sensitive workloads like parsing, extraction, routing, and classification. The source also makes clear that both models support Computer Use and share several baseline capabilities, which may simplify migration planning for teams with mixed workloads.
What to watch
Developers using Gemini APIs will need to check whether their existing requests still include deprecated sampling settings or prefilled model turns. The documentation says those patterns are already ignored or rejected in the current releases and may fail more broadly in future Gemini generations.
It will also be worth watching how teams adapt prompt and output-control strategies now that the older sampling parameters are being phased out for these models. The source suggests Google wants more explicit instructions, structured outputs, and server-side conversation handling instead of legacy request shaping.
Another point to monitor is whether the new defaults and migration guidance change how agents are configured in Gemini Managed Agents. Since Gemini 3.6 Flash now powers the Antigravity agent by default, developers using that environment may need to verify model selection, tuning assumptions, and tool behavior after upgrading.
Finally, the documentation implies a broader direction for the Gemini platform: a more opinionated API surface, tighter validation, and model-specific migration rules. That could reduce ambiguity for production use, but it also means older prompting patterns may become less portable across releases.
