google / gemini-3.6-flash

## Overview Gemini 3.6 Flash is a natively multimodal reasoning model made by Google. It reads text and can also interpret images, audio, and video as input. It generates text outputs that work well for multi-step agent loops, like planning, coding, verifying, and iterating. Its long context window helps it keep large codebases, documents, or long recordings in one conversation. ## What you can build - Coding agents for refactors, bug triage, patch drafts, and test updates - Multi-step workflow runners that plan tasks, execute tool steps, then summarize results - Document and knowledge-work assistants that extract facts, produce drafts, and create checklists - Chart and UI analysis tools that explain visual structure and spot inconsistencies - Meeting and lecture analyzers that summarize long recordings and produce action items - Multimodal QA that answers questions grounded in a mix of text plus attached media ## Inputs - A required instruction in plain text that describes the task and the desired output format - Optional attachments to ground the response in source material, including: - Images such as PNG, JPEG, WebP, HEIC, and HEIF - Videos such as MP4, WebM, QuickTime/MOV, MPEG, WMV, FLV, and 3GPP - Audio such as MP3, WAV, FLAC, AAC, M4A, OGG, PCM, and WebM audio - Optional system-level guidance in plain text that sets rules, style, or role constraints - Optional conversation identifiers (string or number) to keep chat history grouped by user and session - Optional reasoning effort selection with four levels (minimal, low, medium, high) to trade speed for deeper reasoning on hard tasks - Optional sampling controls to change randomness. Some Google deployments ignore custom sampling on this model. - An optional response length cap, up to the model’s maximum output limit ## Outputs Gemini 3.6 Flash returns text. The response can be a normal prose answer, a code patch, a step-by-step plan, or a structured text format like JSON when you ask for it. When you use tool features, the model can produce tool-call style outputs (for example, function arguments) alongside its text, depending on how the run is configured. ## Limitations - The model’s output modality is text only. It does not generate images or audio. - The maximum context window is 1,048,576 input tokens and 65,536 output tokens. - The model can hallucinate details, especially with ambiguous prompts or missing evidence. - Long or noisy media inputs can reduce accuracy. Blurry screenshots, scanned pages, background noise, and low bitrate audio can cause missed details. - Video understanding has practical limits. Some Google documentation notes about 45 minutes for video with audio, and about 1 hour without audio. - Audio understanding also has practical limits. Some Google documentation notes a single audio file per prompt and up to about 8.4 hours, bounded by the 1M-token context. - Some platforms do not support custom sampling knobs (and may ignore them). Some penalty-style controls may be rejected by the API. - The stated knowledge cutoff for Gemini 3.6 Flash is March 2026, so it can be wrong on newer events without grounding. ## Safety & compliance Google provides adjustable safety filters across four categories: harassment, hate speech, sexually explicit content, and dangerous content. Some core harms are blocked by default and can’t be turned off. Usage is governed by Google’s Gemini API terms and prohibited use policies, including restrictions on attempts to bypass safety protections. Google also describes abuse monitoring that can include retention of prompts, contextual input, and outputs for a limited period for policy enforcement.