Gemini 4 Argon: Google Gemini API Shared Context and Model Capability Configuration

One-sentence takeaway

gemini-4-argon shares 1,048,576 context tokens. Synchronous output is capped at 65,536; Long Decode can use the remaining context. Multimodal inputs produce text output.

Use cases

• Suitable tasks: Complex software engineering, deep mathematical and scientific problems, autonomous multi-step Agents, and tasks that require long decoding.

• Unsuitable tasks: Tasks that require audio generation, image generation, or the Live API; tasks that require explicit caching are also unsuitable because the page explicitly states that only implicit caching is supported.

• Applicable model version: Stable version gemini-4-argon.

• Applicable clients, agents, or APIs: Gemini API; the page provides an entry point for trying the model in Google AI Studio. The specific request schema is not shown on this page.

• Recommended reasoning tier and parameters: The page confirms that Thinking supports three tiers: low, medium, and high. The default tier and other sampling parameters are not specified. Use the page's listed output limits for synchronous generation and Long Decode.

Ready-to-use content

The following is the model configuration list published on the page; it does not construct an undisclosed API request body.

Basic identifiers and token limits

Inputs, outputs, and capabilities

Inference and throughput options

The page also marks the following capabilities as Supported:

• Batch API

• Flex inference

• Priority inference

Testing or workflow steps

1. Fix the model ID gemini-4-argon in the Gemini API and confirm that you are using the stable version.

2. Plan the input and output budget: the unified context window is up to 1,048,576 tokens; the maximum output for an ordinary synchronous request is 65,536 tokens; for longer output, use the page's Long Decode mode, which supports up to 1,048,576 tokens.

3. Choose Text, Image, Video, Audio, or PDF according to the task input; the output type is fixed as Text.

4. When reasoning is needed, select and record one of the low, medium, or high tiers. The page does not specify a default tier, so do not assume one.

5. If caching is used, confirm that the workflow depends on implicit caching. Do not make explicit caching part of the acceptance criteria because the page marks Explicit caching as Not supported.

6. Verify Code execution, Computer use (Preview), File search, Function calling, Google Maps grounding, Search grounding, Structured outputs, and URL context individually. Do not treat Supported as proof that the capability is enabled for the target account or environment.

7. Test unsupported paths: Audio generation, Image generation, and Live API should be marked Not supported; for Agentic video understanding, also record the default static mode.

8. Record actual input/output tokens, synchronous or Long Decode mode, Thinking tier, latency, and failure reason. The page does not provide a complete API schema, quotas, or sampling parameters, so they cannot be filled in from this page.

Raw evidence and data

• The page defines Gemini 4 Argon as a flagship reasoning model for complex software engineering, deep mathematical and scientific problems, autonomous multi-step Agents, and Long Decode of up to 1M tokens.

• The model code listed on the page is gemini-4-argon, and the stable version is also gemini-4-argon.

• The page defines Context window as the unified input-and-output token limit, with a value of 1,048,576.

• The page lists Max output as 65,536 for default synchronous generation and up to 1,048,576 with Long Decode.

• The listed input types are Text, Image, Video, Audio, and PDF; the output type is Text.

• Caching is marked Supported (Implicit only), with the explicit note that Explicit caching is not supported.

• Thinking is marked Supported, with tiers low, medium, and high.

• Agentic video understanding is marked Supported (defaults to static), and Computer use is marked Supported (Preview).

• Code execution, File search, Function calling, Grounding with Google Maps, Search grounding, Structured outputs, URL context, Batch API, Flex inference, and Priority inference are marked Supported.

• Audio generation, Image generation, and Live API are marked Not supported.

• The page provides a Google AI Studio trial entry point, Long Decode documentation, the Latest model page, and the Gemini 4 Argon model card.

Scope and limitations

• 1,048,576 is the unified input-and-output token limit defined by the page; do not interpret it as an additional output budget of the same size on top of the input limit.

• 65,536 is the default synchronous maximum output; 1,048,576 is the Long Decode maximum output. They correspond to different generation modes.

• The page only states that Thinking supports low, medium, and high; it does not specify a default tier, the cost of each tier, or quality differences between them.

• “Supported” only means that the model page lists the capability. It does not mean that the capability is enabled for the current account, client, region, or specific request; verify it separately in the target environment.

• Caching is implicit only; explicit caching cannot be a prerequisite for this model's configuration.

• The page does not provide a complete API request schema, authentication method, quotas, pricing, regional availability, sampling parameters, or detailed limits for each capability. Consult the relevant documentation rather than inferring these fields from this page.

• Computer use is in Preview, and Agentic video understanding defaults to static. Retain version and behavior-change risk for both capabilities.

Reproduction notes

1. Open the Gemini 4 Argon model page and confirm the stable version and last-updated date shown on the page.

2. Run a fixed set of inputs with gemini-4-argon, recording the actual results for Text, Image, Video, Audio, and PDF inputs and Text output.

3. Verify the 65,536 maximum output in synchronous mode separately from the page limit of up to 1,048,576 in a Long Decode workflow; do not conflate the two modes.

4. Record the actual configuration for the low, medium, and high Thinking tiers. If the page does not specify a default, record it as “Not specified.”

5. Verify the implicit caching path, and confirm that explicit caching, Audio generation, Image generation, and the Live API are not available capabilities.

6. Verify each capability marked Supported by the page in the target environment, saving the actual request configuration, token usage, latency, and errors. Do not fill in an API schema that the page does not publish.