Loading model details…
Google's flagship multimodal reasoning model for advanced software engineering, math, science, long-running agents, and professional work with a unified 1M-token window.
| Item | Specification |
|---|---|
| Input | Text, image, video, audio, PDF |
| Output | Text |
| Context | 1,048,576 tokens shared by input, thinking, and output |
| Max Output | 65,536 tokens for default synchronous use; Long Decode can use the remaining unified window |
| Vision | ✓ Supported |
| Function Calling | ✓ Supported |
Gemini 4 Argon is Google's generally available flagship reasoning model released broadly in October 2026 under the model code gemini-4-argon. It accepts text, image, video, audio, and PDF inputs and produces text within a unified 1,048,576-token window shared by the input, thinking, and visible output.
The model targets complex software engineering, mathematical and scientific reasoning, autonomous multi-step agents, enterprise knowledge work, and cyber defense. Standard synchronous requests generate up to 65,536 tokens by default; background Long Decode continues across multiple stages and can use up to 1,048,576 tokens minus input and thinking. It supports low, medium, and high thinking with medium as the default, plus code execution, File Search, function calling, Google Search and Maps grounding, structured output, URL Context, and Preview Computer Use.
| Feature | Benefit |
|---|---|
| Unified 1,048,576-Token Window | Input, thinking, and output share one allowance for large sources and long results. |
| Long Decode | Background execution crosses the synchronous 64K boundary through continuation stages approaching the unified limit. |
| Three Thinking Levels | low, default medium, and high cover light analysis, general agents, and the hardest reasoning. |
| Complete Tool Ecosystem | Code, files, functions, Search, Maps, and URL Context enable verifiable workflows. |
| Five Native Input Types | Text, images, video, audio, and PDFs support cross-media research and agents. |
| Stable Production Version | A fixed GA model code with Batch, Flex, and Priority inference supports deployment. |
The 1,048,576-token limit is shared by the input, thinking, and final output rather than assigned separately to each. Longer prompts or deeper reasoning leave less room for visible output.
Long Decode lets a background task continue through multiple stages of up to 64K tokens each, producing more than 65,536 tokens within the remaining unified window. Applications enable background execution, raise max_output_tokens, and join all model_output steps after completion.
Use low for light analysis and simple writing, the default medium for most complex coding and agent work, and high for difficult mathematics, science, and the hardest code tasks. Evaluate correctness, first-pass completion, response time, and tool loops on representative workloads.
No. It supports implicit caching only, so migrations should remove cachedContents, caches.create, and explicit cached_content references and then retest long-prompt workflows.
Change the model ID to gemini-4-argon, review max_output_tokens and thinking levels, and remove ignored temperature, top_p, and top_k settings. Use background execution for output beyond 64K, increase timeouts, and preserve server-side conversation state for multi-turn tools.
Gemini 4 Argon is available on LinkModel. Use the public model ID, request schema, context limits, and thinking controls shown on LinkModel's live page, testing standard responses before multimodal input, tools, and long output.
Test short text, long context, image or PDF input, structured output, function calling, and refusal handling separately, and confirm the effective unified-window limit. Do not assume Google's Interactions API, background Long Decode, implicit caching, or provider-specific tool fields pass through unchanged.