Use Gemini 4 Argon for text reasoning, multimodal understanding, structured output, function calling, and grounded tool workflows.
Gemini 4 Argon is Google’s flagship reasoning model for complex software engineering, deep mathematical and scientific problem solving, autonomous agents, and long-running multimodal workflows. Use the stable model ID gemini-4-argon with POST /v1beta/models/gemini-4-argon:generateContent.
Long-context reasoning — Share up to 1,048,576 input, thinking, and output tokens in one context window.
Multimodal understanding — Analyze text, images, video, audio, and PDF documents; the model returns text.
Agentic workflows — Use function calling, code execution, Google Search, Google Maps, URL context, and Computer Use (Preview).
Structured output — Constrain responses with responseMimeType and responseSchema.
Long decode — Google GenAI SDKs can continue responses beyond the per-request cap. AnyFast’s native endpoint documents the synchronous generateContent request shape shown below.
Send text in contents[].parts[].text. Use generationConfig.thinkingConfig.thinkingLevel to trade latency for reasoning depth.
Level
Recommended use
low
Latency-sensitive chat, drafting, and quick analysis
medium
General coding and agent workflows; this is the default
high
Difficult reasoning and multi-step tool orchestration
The final conversation turn must be a non-empty user turn. Do not prefill a response by ending the request with a model turn.
Combine text with images, audio, video, or PDF documents in the same contents request.
Use inlineData with mimeType and Base64-encoded data for inline media.
Use fileData with mimeType and fileUri for an accessible media or document URI.
Keep each media item inside the parts array for the turn that refers to it.
The model returns text; it does not generate image, audio, or video output.
Base64 encoding increases request size. Prefer fileData for large assets or files reused across requests.
Set generationConfig.responseMimeType to application/json and supply generationConfig.responseSchema to constrain the output.Use a focused schema with explicit required fields. The returned JSON is contained in a text part, so parse candidates[].content.parts[].text after checking finishReason.
Add functions under tools[].functionDeclarations. Control selection with toolConfig.functionCallingConfig.Supported modes are AUTO, ANY, NONE, and VALIDATED. Use allowedFunctionNames to restrict the callable subset.Preserve every thoughtSignature returned by the model when replaying conversation history. A functionResponse must include the function name; return the model-provided call id and your structured response when available.
Gemini 4 Argon can use the following built-in tools on the current endpoint:
Tool field
Purpose
googleSearch
Ground answers with current web results
googleMaps
Ground location answers with Google Maps data
urlContext
Read and analyze supplied web pages
codeExecution
Run model-generated code and return its result
computerUse
Propose browser actions; this capability is in Preview
Read the returned groundingMetadata, executableCode, codeExecutionResult, or tool-call parts instead of assuming every response contains only text.
Use user and model roles for conversation history.
Preserve complete model turns and every thoughtSignature in multi-step function workflows.
Check candidates[].finishReason before consuming the output.
Read token counts from usageMetadata, including thoughtsTokenCount for reasoning usage.
Search and Maps responses can include groundingMetadata with queries, sources, and supporting segments.
Code execution can return executableCode and codeExecutionResult parts.
Do not send temperature, topP, topK, or candidateCount. Google documents these sampling controls as ignored for Gemini 4 Argon. Explicit context caching is also unsupported; implicit caching may apply automatically.
API Reference
View request fields, capability examples, and response schemas for Gemini 4 Argon.
