Google: Gemini 3.6 Flash (batch)

Gemini 3.6 Flash is Google’s multimodal workhorse profile for coding, knowledge work and agentic execution. It builds on Gemini 3.5 Flash, with a focus on improved multimodal performance and token efficiency. The model accepts text, images, audio, video and PDF and returns text. A useful assignment can combine code, a screenshot and a requirement, or compare a chart or visual plan with written constraints to produce a checked explanation or targeted revision.

The developer guide emphasizes fewer unnecessary reasoning steps, turns and tool calls in multi-step work, along with more focused edits. For diagnosis, provide the observed failure, the relevant source and the behaviour that should remain unchanged. For chart debugging, ask the model to inspect why a chart renders the wrong series, identify the data transformation involved and propose one correction while retaining the surrounding layout. The result can then be compared with the actual input data and the required chart behaviour.

Upfront programmatic inspection is a distinctive behaviour described for 3.6. The maker says the model more often inspects a problem with diagnostic code before making changes. This can help a complex investigation, but the guide notes that it can add exploratory steps to routine frontend work. State whether the task needs an investigation, a simple edit or a design comparison. Supply the information that makes that distinction clear and specify the permitted actions for the connected application’s tools.

Spatial and multimodal reasoning are also highlighted: chart interpretation, visual blueprint conversion and multi-element web layouts. Supply the relevant view or pages, identify what should be interpreted and ask for the evidence behind the conclusion. For a design task, include concrete layout and styling guidelines. The maker notes a functional-code emphasis and that evaluators sometimes preferred earlier models’ visual styling, making an explicit visual brief useful. Check the result against the actual diagram or design requirements.

Medium is the documented default thinking level. Google’s model-specific thinking table lists minimal, low, medium and high; the 3.6 migration guide uses medium or high in its effort examples. The exact page lists function calling, code execution, file search, search grounding, Maps grounding, URL context and structured outputs, with Computer Use as a preview capability. Define function inputs and the required result clearly. A function selected by the model needs a real application result before a later answer can rely on its outcome. For a structured analysis, define the schema and source evidence needed for each finding.

The 3.6 guide changes preparation compared with earlier Gemini requests. It deprecates temperature, top-p and top-k sampling settings, rejects a prefilled final model-role turn and recommends system instructions or structured outputs for controlling the answer. Its function-response rules match a response to its call by identity and name, and the Gemini 3 baseline preserves thought signatures in multi-turn use. A practical continuation should carry the unmodified relevant history and tool result rather than imitate a previous answer prefix.

The native card specifies up to 1M input context and 64K text output; the exact API table gives 1,048,576 input tokens and 65,536 output tokens. Budget the instructions, history and media as part of the input, and ask for an output sized for the task. The producer page also lists caching, batch, Flex and Priority processing. When processing independent records, keep the record identifier and result format explicit. The output is a text artifact, even when the evidence includes photographs, audio or video.

Possible limitations include hallucinations and occasional slowness or timeouts. Its March 2026 cutoff includes domains whose knowledge remains limited to January 2025, so attach newer material or use a documented grounding workflow when current facts matter. Review calculations, source references, visual interpretations and proposed modifications. The distinctive preparation for this edition is a focused diagnostic brief, clear tool scope and explicit design constraints, followed by inspection of the concrete result.