Google: Gemini 3.5 Flash (batch)

Gemini 3.5 Flash is Google’s multimodal Flash model profile for sustained coding and agentic work. Its documented inputs are text, images, video, audio and PDF, with text output. The model’s task focus includes iterative coding, exploration of alternate solutions and longer sequences of instructions and tool use. The 3.5 edition provides documented thinking controls, multimodal inputs and tool workflows.

A coding task can combine a written requirement, relevant source files and a screenshot of the intended result. Supply the observed behaviour, the expected behaviour and the constraints that must stay intact. An illustrative brief is: “Compare the screenshot with the component text, identify the mismatch in the form workflow and propose one focused correction. Explain which source details justify the change.” For a document task, provide the relevant pages and ask for a comparison against explicit requirements, with source references for each finding.

The edition documents minimal, low, medium and high thinking levels. Medium is the default, changed from the high default of Gemini 3 Flash Preview. Minimal suits simple answers and straightforward tool decisions without guaranteeing that reasoning is completely disabled; low targets tasks requiring fewer steps; high provides deeper reasoning for hard mathematics, coding and agent problems. Google recommends using thinking_level for effort control and retaining default sampling settings rather than trying to force reasoning by changing temperature, top-p or top-k.

A defining behaviour is preservation of reasoning context across a multi-turn task. Google’s Interactions workflow carries that context automatically; its GenerateContent workflow requires the full unmodified conversation history, including thought signatures. This supports a continuing debugging or refactoring assignment in which earlier reasoning informs the next step. Preserve the actual conversation structure rather than converting each follow-up into an unrelated prompt, and keep the final requested artifact clear as the work progresses.

The documented tool set includes Google Search, grounding with Google Maps, File Search, Code Execution, URL Context and standard Function Calling, with combined tool use and Computer Use marked as a preview capability. Structured Outputs can supply a defined data record. Give functions a clear purpose and parameter definition: the model chooses a function, while the connected application performs its operation and returns the result. Google requires function results to match the preceding calls by identity, name and count; multimodal results belong inside the function response.

The exact API model table gives a maximum input of 1,048,576 tokens and maximum output of 65,536 tokens. Conversation history and supplied media contribute to the input budget. Caching and batch processing are documented for this profile, with additional Flex and Priority processing options in Google’s table. These limits and capabilities describe the producer edition; they do not require an answer to fill the entire output window or make every tool necessary for a task.

This profile generates text rather than images or audio and does not use Live API. Its developer guide also excludes image segmentation. For a visual question, ask for a description, extraction or interpretation of the supplied material; for a spoken recording, ask for a text result. When the answer depends on newer information, the documented search-grounding workflow can add sources beyond training knowledge. State whether external information is needed and what evidence the final answer should retain.

For a useful longer assignment, define the final result and the completion criterion before the first step. A support-analysis task might request a table of observed issues, supporting passages and a proposed reply; a coding task might require the focused change and the relevant validation result. Attach the source context that matters, specify the permitted tool actions and request an explicit final answer that separates completed work from questions still requiring information. This turns the model’s sustained workflow into a concrete text, code or structured artifact.