Gemini 4 Argon is Google's first new-generation Gemini model. It was showcased to a select group of cybersecurity specialists on September 30, 2026, and made available on October 13 to subscribers of Google AI Pro and Ultra within the Gemini app and AI Mode, as well as to developers via the paid tier of Gemini API. Below is a practical breakdown: what has changed compared to Gemini 3.x, how it works from an API perspective, where the model can be useful for developers and webmasters, and how to rewrite prompts. The material is compiled from official Google documentation and announcements. I will add my own tests separately where they are marked.
⚡ In Brief: What's Important to Know About Gemini 4 Argon
- ✅ Unified Window of 1,048,576 Tokens. According to the Gemini API documentation, input tokens, thinking tokens, and output tokens share a single 1M token window.
- ✅ Up to 1M Tokens Output (Long Decode). Google reports that the output limit has increased from 64K to 1M tokens.
- ✅ Three Thinking Levels:
low, medium (default), and high.
- ⚠️ This is Not a Free Model. It's only available in the API on the paid tier and in the app for AI Pro and Ultra subscribers. The introductory price is valid until November 30, 2026, after which it doubles.
- ⚠️ Most Figures Are Self-Reported by Google. Benchmarks and statements about response style have not yet been confirmed by independent measurements.
📚 Table of Contents
🕰️ How It Was: Gemini 3.x and Its Limitations
The previous generation, Gemini 3.x: Gemini 3 Pro was released on November 18, 2025, followed by 3.1 Pro and the Flash line (the latest at the time of writing being 3.8 Flash). Here are the bottlenecks, according to Google's documentation and statements.
Generation Limits
Google states that the new output limit of 1M tokens has increased from the previous 64K. For long tasks (large migrations, voluminous reports), this meant manually breaking down work into steps and stitching the results together.
Thinking Control
Previously, thinking costs were regulated by the thinking_budget parameter. In Gemini 4 Argon, the primary regulator is thinking_level. According to the migration checklist, thinking_budget still works and maps to the nearest level, but returns a warning.
Response Style
In its announcement about the launch in the Gemini app, Google describes Argon as responding more directly and with fewer caveats (hedging), not repeating itself unnecessarily, and avoiding clichés. I haven't seen independent measurements of this, so this is a company claim, not a verified fact.
Agency: What Was and What's New
The assertion that "Gemini 3 only answers questions" would be inaccurate. Along with Gemini 3, Google released Antigravity, a platform for developing with agents, and in May 2026, the background agent Gemini Spark based on Gemini 3.5 Flash appeared. What's new in Argon is the claim of sustained performance on long, multi-step tasks ("long-horizon"), not the concept of agents itself.
🧩 What's New in Gemini 4 Argon: Key Updates
Google does not publish architectural details, so below are product and API changes that can be verified in the documentation.
Unified 1M Token Window
The 1,048,576 token window is shared for input, thinking, and output. The practical consequence is that the more the model thinks and writes, the less space is left for your data. Large responses at each step of the dialogue quickly fill this window.
Long Decode: Output Beyond a Single Request Limit
| Parameter |
Value According to Documentation |
| Single Streaming Request Limit |
65,536 tokens |
| Single Non-Streaming Request Limit |
32,768 tokens (was 65,536 on Gemini 3.x) |
| Overall Limit with Long Decode |
Up to 1M tokens minus your input |
| Automatic Continuation |
Enabled by default in SDKs: Python google-genai from 2.30.1, JavaScript @google/genai from 2.29.1 |
If you are using REST or an older SDK, the request stops at the single request limit, and you need to implement continuation yourself (there's a continuation_token). On a complex task, the model might spend its entire single "hop" on thinking and return empty text with a continuation_token. The documentation calls this normal behavior.
Managed Thinking Levels
| Level |
Token Consumption |
When to Use (According to Documentation) |
low |
Lowest, lowest latency |
Simple tasks and cost savings. May be suitable for analysis and writing that only requires light thinking; quality may decrease on complex tasks. |
medium (default) |
Balanced |
Most tasks, including complex code and agent scenarios. |
high |
Highest, slowest first token |
Complex reasoning, math, science, most challenging code and agent tasks; pairs well with long decode. |
Thinking tokens are counted within max_output_tokens and are billed as output tokens.
Agent Capabilities and Tools
According to the model page, Argon supports function calling, code execution, Google Search and Maps integration, URL context, File Search, and structured responses. Computer use is available in Preview mode. The Antigravity agent in Gemini Managed Agents also supports Argon. The Gemini app has introduced a "Tasks" mode which, according to Google's description, can run while you're away from your computer, within Google apps, and across 40+ connected services.
Multimodality: What's Included and What's Not
- Input: text, images, video, audio, PDF.
- Output: text only. This model does not have image generation, audio generation, or Live API capabilities.
- Agentic video understanding is now generally available (GA).
Tone
Direct responses are Google's claim for the Gemini app. There are no such formulations in the API documentation, so in the prompts section, I suggest testing this with your own tests.
Pricing and Access
| Item |
Conditions |
| Introductory Price (until November 30, 2026) |
$2 per 1M input tokens and $10 per 1M output tokens (including thinking tokens) |
| From December 1, 2026 |
$4 per 1M input tokens and $20 per 1M output tokens |
| Caching |
Implicit caching is enabled for requests over 4,096 tokens and reduces the price of cached input tokens by 95%. Explicit caching (cachedContents) is not supported. |
| API |
Paid tier of Gemini API. According to the documentation, the model is not available on the free tier or in the Google AI Pro plan. |
| Gemini App and AI Mode |
For Google AI Pro and Ultra subscribers. |
Sources: API documentation, app announcement. Prices and conditions are subject to change, so please refer to the Pricing page before launching.
⚙️ How the Model Works: Processing Complex Tasks
Google does not disclose its internal architecture. Below is the sequence visible from the API documentation. This describes what happens at the request level, not a neural network diagram.
- Request and Context. Your prompt, files, and dialogue history are fed into a shared window of 1M tokens (read my article on LLM context windows for details on how model memory works and why context limits affect cost and quality).
- Thinking Phase. The model consumes thinking tokens based on
thinking_level. These count towards your output limit and are billed as output. In multi-turn dialogues and function calls, you need to pass back the full history along with thought signatures.
- Tool Calling. The model uses built-in tools (code execution, search, URL context) within the scope of the request. For function calling, the situation is different: the model returns a call, you execute it in your code, and return the result. Thus, "human-free" in the API means "human-free in the loop," but the developer writes the loop.
- Final Response. If the request exhausts
max_output_tokens, the response is returned with finish_reason: "MAX_TOKENS" and a continuation_token. Newer SDK versions continue automatically, while REST requires a custom loop.
Example Request (Python)
pip install -U "google-genai>=2.30.1"
from google import genai
from google.genai import types
client = genai.Client()
response = client.models.generate_content(
model="gemini-4-argon",
contents="Describe a plan for migrating a module from Java 8 to Java 21.",
config=types.GenerateContentConfig(
max_output_tokens=65536,
thinking_config=types.ThinkingConfig(thinking_level="medium"),
),
)
print(response.text)
The example is based on the official quickstart. Check parameter names and minimum SDK versions before use.
What Changed in the API (Migration Checklist)
| Change |
Action Required |
thinking_budget |
Replace with thinking_level; test with low, medium, high |
| Output Limits |
Set max_output_tokens explicitly; raise values below 32K to 32,768 |
| Sampling Parameters |
Remove temperature, top_p, top_k, candidate_count: they are ignored |
| Explicit Caching |
Not supported and returns an error; use implicit caching |
| Timeouts |
Set client HTTP timeout to at least 20 minutes per call |
| Thought Signatures |
Pass the full history with signatures in multi-turn requests |
🛠️ Where and How to Use: Practical Scenarios
Below are two types of statements: Google's claims (which are worth verifying) and my assumptions about application (marked as such).
Development and Refactoring
- Code Migrations. Google provides an example: Argon agents rewrite C/C++ to Rust, from tens of thousands of lines in core libraries to over 800,000 lines for the Zircon kernel in Fuchsia. This work, according to Google, undergoes automatic and manual audits, emulation testing, and review before production release. This process should be adopted as a rule for your own projects.
- Benchmarks. On DeepSWE v1.1, which measures long, real-world development tasks, Google claims 77.9% and calls it a new state-of-the-art level. This is data from the company itself.
- Security. Google positions the ability to find and fix vulnerabilities for defenders. It provides a version with relaxed restrictions only to verified participants of the Fairwind program, so do not rely on it in public access.
Analytics and Data Handling
- Large Documents. The shared 1M token window allows you to pass extensive documentation or reports in a single request, but don't forget that the response and thinking also consume space. Calculate tokens in advance using token counting.
- Server Logs (My Assumption). Similar analysis is possible, but before sending logs to an external service, remove or anonymize IP addresses and other personal data, and check the data usage policies of your tariff.
- Finance and Law. Google claims leadership on the Vals Index, Vals Finance Agent, and Harvey Legal Agent Benchmark. This is a self-report, so always manually verify critical conclusions.
Content and SEO (My Assumptions)
Google does not publish separate claims about SEO scenarios, so this section contains my ideas for practical testing:
- Building semantic structures for articles and clusters based on your data (exports from Search Console, list of queries).
- Technical audit based on crawler exports: grouping errors, prioritization, draft fixes.
- Preparing drafts of analytical materials with a summary in the form of fact-checking: verify all numbers and links against primary sources.
For publishers, there's another detail: according to a report on Search Engine Roundtable, Argon's responses in AI Mode include links to sources. This is worth monitoring in analytics, but the conclusions themselves should still be considered preliminary.
✅ Checklist: How to Craft Prompts for the New Model
1. Choose the Thinking Level Based on Task Type
low — classification, rewriting, simple answers, mass low-cost requests.medium — typical choice, including for code and agents.high — complex reasoning, most difficult code, long outputs.
Google recommends running your evaluations at all three levels with a high max_output_tokens and selecting the lowest level that meets your quality bar.
2. Are Preambles Really No Longer Needed?
In the Gemini app, Google claims direct answers without clichés. I haven't found such a guarantee in the API documentation. Try it yourself: run the same prompt with the phrase "answer directly, without fluff" and without it, and compare. If there's no difference, remove the preamble. A clear structure works better than general wishes (and you can find ready-made templates and examples for content and development in my collection of practical prompts for code and text generation):
Goal: Perform a technical audit of a URL list from a crawler export.
Context: [insert data].
Constraints: only facts from the data, no assumptions; if data is missing, state what is missing.
Format: table "issue | number of URLs | priority | what to fix".
Verification: at the end, list what you could not verify.
2. Are Preambles Really No Longer Needed?
In the Gemini app, Google claims direct answers without clichés. I haven't found such a guarantee in the API documentation. Try it yourself: run the same prompt with the phrase "answer directly, without fluff" and without it, and compare. If there's no difference, remove the preamble. A clear structure is more valuable than general wishes: goal, context, constraints, output format, verification criteria.
Goal: Perform a technical audit of a URL list from a crawler export.
Context: [insert data].
Constraints: only facts from the data, no assumptions; if data is missing, state what is missing.
Format: table "issue | number of URLs | priority | what to fix".
Verification: at the end, list what you could not verify.
3. How to Transfer Large Files and Codebases
- Calculate tokens before sending and leave a buffer: thinking and response also take up space from the same 1M.
- For files, use the Files API rather than inserting everything into the prompt when appropriate.
- Set
max_output_tokens explicitly. Guidelines from the documentation: for standard tasks, leave the default value; for tasks that previously required over 32K tokens, set it to 65,536; for complex reasoning and code, set it to 131,072 or 262,144, up to 1,048,576 minus input.
- For savings, use implicit caching: it works for requests from 4,096 tokens and reduces the cost of cached input tokens by 95%.
4. Control Costs
- Thinking tokens are billed as output, and output at $10 (then $20) per 1M is noticeably more expensive than input.
- For long tasks, limit
max_output_tokens and log actual token usage.
- Remember the tariff changes effective December 1, 2026.
❓ Frequently Asked Questions (FAQ)
Are Gemini 4 and Gemini 4 Argon the same?
Argon is the first model in the Gemini 4 family. According to press reports, Google plans to introduce the rest of the family later.
Is Gemini 4 Argon free?
No. The model is available on a paid tier in the API and, according to the documentation, is not available on the free tier or in the Google AI Pro plan. In the Gemini app, it is available to AI Pro and Ultra subscribers.
How much does it cost to use?
Until November 30, 2026, the introductory price is $2 per 1M input and $10 per 1M output tokens. From December 1, it will be $4 and $20. Thinking tokens are billed as output.
What is long decode?
This is the ability to generate a response longer than the single-request limit (64K in streaming, 32K non-streaming), up to 1M tokens minus the input. Newer SDK versions continue generation automatically.
Can thinking be disabled entirely?
The documentation lists tiers as low, medium, and high. There is no "disabled" mode among them. The cheapest option is low.
🏁 Conclusions
- 🔹 Gemini 4 Argon is the first model of the Gemini 4 generation. It became publicly available on October 13, 2026, after being available only to cybersecurity partners for several weeks prior.
- 🔹 Key practical changes include: a unified window of 1M tokens, long decode up to 1M output tokens, and
low / medium / high thinking levels instead of thinking_budget.
- 🔹 When migrating, remember the nuances: sampling parameters are ignored, there is no explicit caching, the single non-streaming request limit has decreased to 32K, and timeouts need to be increased.
- 🔹 Claims about response style and benchmarks are self-reported by Google. Before switching to a new model, test it with your own data.
- 🔹 The most cost-effective strategy is to test tasks on all three thinking levels and choose the cheapest one that meets your quality requirements.
📎 Sources
Important: Gemini models, limits, pricing, and access conditions are subject to change. Please check the current documentation before use. This material is current as of October 18, 2026.
