Google’s flagship reasoning model for complex coding and analysis, with a shared 1M-token context, multimodal understanding and provider long-decode support.
65,536 synchronous default; up to 1M minus input with long-decode continuation
Inputs → output
Text + Images + Video + Audio + PDF → Text
Knowledge cutoff
Not verified
API model ID: gemini-4-argon
Capabilities & boundaries
What it supports. Where the limits are.
Tool support requires the appropriate API integration; a supported tool is not automatically active in every chat.
Supported API features and tools
A unified 1,048,576-token window shared by input, thinking and output, rather than independent million-token input and output budgets
A default synchronous output limit of 65,536 tokens, with longer responses through the provider’s documented long-decode mechanisms
Low, medium and high thinking levels on the provider API, with medium as the documented default
Provider multimodal understanding, Google Search grounding and compatible tool-driven workflows
Implicit context caching; Google does not support explicit context-cache creation for Argon
Before you choose
The 1,048,576-token window is shared by input, thinking and output. A larger source pack or reasoning allocation reduces the remaining space for the visible response.
Long decode needs background execution or a supported continuation loop, plus enough shared context for the input and output. The synchronous default and long-decode ceiling are different provider limits.
Argon understands multimodal input but produces text. Use a dedicated image, video or audio generation model when the deliverable is new media.
Google’s Argon guidance says supplied temperature, top_p and top_k controls are ignored. Use the documented thinking levels and output budgets rather than relying on those sampling settings.
A large context or longer reasoning does not guarantee accurate citations, successful code or complete recall. Keep output bounded and verify source claims and proposed changes independently.
Gemini thinking levels
lowmedium · defaulthigh
Google supports low, medium and high thinking, with medium as the documented API default. Start with medium for a representative task, compare low for simpler requests and high for work with interacting constraints. Thinking level manages reasoning expenditure and latency; it is not a fixed visible-answer length or a guarantee of correctness. The output budget covers both thinking and the visible response.
The stable model ID is gemini-4-argon. Google’s Gemini Developer API provides GenerateContent and streamGenerateContent for its native content-part and generation-configuration format.
Set thinkingConfig and the output-token budget deliberately for the task. Google’s migration guidance recommends reviewing existing small budgets when moving to Argon; input, thinking and the final answer all consume shared context.
Google Search grounding, function calling and structured outputs have different request contracts. Configure the intended capability, return tool results where needed and validate citations or structured responses against their source requirements.
Google’s long-decode documentation distinguishes the synchronous default from larger responses using background execution or client-side continuation. Input, thinking and output still share one 1,048,576-token window.
Put it to work
Start with a more useful prompt.
Original examples from EZ Ai Assist. Adapt these to your task and the features available in your workspace.
Workflow 01
Analyze a payment retry design
Reason about pasted code without implying the model executed it.
Review the payment retry flow and code excerpts below. Trace how duplicate charges or lost acknowledgments could occur across a timeout, restart and repeated webhook. Cite the supporting code for each scenario. Propose idempotency boundaries, a minimal test matrix and a staged repair, keeping assumptions explicit. Do not claim you executed transactions or verified a production payment system.
Workflow 02
Build an evidence-based research brief
Use supplied passages, with optional app search for current sources.
Using the dated source passages I paste below, compare the two technical approaches against the stated requirements. For each recommendation, quote a supporting passage, identify conflicting evidence and explain what additional measurement would change your conclusion. Keep the answer under 1,200 words. If web search is enabled, distinguish newly retrieved sources from my supplied material and cite each current claim.
Workflow 03
Review a technical diagram
Check visual evidence against a short written specification.
Inspect this architecture diagram and the requirements pasted below. Trace the visible data path, list components whose role is unclear, and identify potential single points of failure supported by the diagram. Separate missing information from an actual design defect. Suggest a compact verification checklist and a revised component description, without inventing connections or treating a drawn arrow as proof of a working integration.
Developer reference
Google API pricing
These are published provider API reference prices, not EZ Ai Assist subscription prices or verified customer charges.
Google Gemini Developer API · Standard paid tier · published date schedule · USD per 1,000,000 tokens
Token type
Price
Input · through November 30, 2026
$2.00
Cached input · through November 30, 2026
$0.10
Output · through November 30, 2026
$10.00
Input · from December 1, 2026
$4.00
Cached input · from December 1, 2026
$0.20
Output · from December 1, 2026
$20.00
Rates checked October 16, 2026. The introductory period ends November 30, 2026; the second set of rows shows Google’s published rates starting December 1, 2026. Both periods are labeled explicitly rather than silently switching a timeless price.
Output prices include thinking tokens. Implicit caching discounts eligible cached input by 95%; explicit context caching is not supported. Google Search grounding can add separate usage charges under the provider’s current pricing policy.
Google lists no free API tier for Argon. Google developer access and consumer subscriptions are separate from EZ Ai Assist account eligibility and plan allowances.
These dated Google developer rates are separate from EZ Ai Assist plans and usage charges. Use the app’s plan and usage information for customer-facing costs.
Common questions
A few things worth knowing.
When should I choose Gemini 4 Argon?
Try it for a difficult coding, engineering or analysis brief with explicit evidence and acceptance criteria. Keep the deliverable bounded and compare its result with another model on your own task. Its flagship positioning does not guarantee correctness or remove the need to test suggested code.
Is the context 1M input plus 1M output?
No. Google documents one unified 1,048,576-token budget shared by input, thinking and output. Conversation history and source material use part of that space, so the maximum possible response shrinks as the input grows.
How does long decode differ from an ordinary response?
Google documents a 65,536-token synchronous default. Larger responses use background execution or client-side continuation and remain bounded by the shared 1,048,576-token input, thinking and output window. This is not a separate million-token output allocation on top of a full input window.
How do Argon’s thinking levels work?
Google supports low, medium and high, defaulting to medium. Compare lower effort for straightforward tasks with higher effort for complex dependencies. Thinking tokens consume output budget and are billed at the output rate; a higher level does not guarantee a more accurate answer.
What documents and media does Argon understand?
Google documents text, images, video, audio and PDF inputs with text output. Ask for page, timestamp or passage references and verify important details against the source. Understanding supplied media is different from generating a new image, video or audio file.
What does Google Search grounding add?
Grounding retrieves current information through Google Search and supplies source information for the answer. Check the returned sources and confirm time-sensitive claims. Function calling and structured outputs are separate API capabilities with their own configuration and validation needs.
Why are two sets of prices shown?
Google publishes introductory rates through November 30, 2026 and higher rates starting December 1, 2026. Each row names its period, and output rates include thinking tokens. These are developer API references; your EZ Ai Assist plan and usage accounting are separate.
Can I tune Argon with temperature or top_p?
Google’s Argon migration guidance says temperature, top_p and top_k values are ignored. Use the documented thinking controls, a clear task brief and an intentional output budget instead of assuming those sampling settings change the result.
Check the source
Official documentation
Specifications and API prices checked on . Example prompts and workflow advice are editorial guidance from EZ Ai Assist.