Gemini 4 Argon: What Google’s New Flagship Can Do

From Closed Launch to General Availability

Argon rolled out in two stages. On September 30, Google announced the model and granted access to vetted cybersecurity professionals via the Fairwind program. The company explained this sequence on security grounds: releasing capabilities at this level requires a phased approach. In parallel, Google took part in a voluntary U.S. government process that provides access to models ahead of release.

Fairwind launched in early September alongside Gemini 3.8 Flash Cyber — a version designed for vulnerability discovery.

Gemini 3.8 Flash: What’s New in Google’s Third Model in a Month and a Half and How to Build a 3D Game

04.09.2026

Read now

Program participants get Argon without “cyber restrictions” so defenders can tap the model’s full potential.

On October 13, the second stage began. The model became available via the API through AI Studio and the Gemini platform, in Antigravity, the Gemini app, and AI Mode in Google Search.

The two-week gap between the announcement and the release was used to refine safety mechanisms. In a September post, Google said it would collect feedback from early testers and only then open the model to developers, companies, and everyday users.

What Is Gemini 4 Argon

According to the model card, Gemini 4 is the next generation in the lineup, and Argon is “a new frontier model, not a modification or fine-tuning of the previous one.”

Inputs. Text, images, video, audio, and PDFs. Output is text only — Argon cannot generate images or audio.

Context. A single window of 1,048,576 tokens for input and output combined.

Response length. By default — up to 65,536 tokens. In long decode mode — up to 1,048,576 tokens.

Reasoning levels. Low, medium, and high.

Tools. Function calling, code execution, file search, grounding in Google Search and Google Maps, structured output, URL context, and computer control in preview mode.

Knowledge cutoff. March 2026, but in some areas the model’s knowledge may be less up to date.

The main innovation is the response limit. Google calls it “best in class” and explains it like this: if a model has headroom to generate hundreds of thousands of tokens in a single trajectory, it can reason more deeply and solve complex tasks in one pass. Previous Gemini models were capped at 64,000 tokens.

Google also highlights the response style. According to the company’s statement, Argon answers more directly, hedges less often with caveats, avoids repetition, and steers clear of clichés. On long tasks, the model more often reports on its progress.

Benchmarks: Where Argon Leads and Where it Falls Behind

Google compares Argon with three competitors: OpenAI’s GPT-6 Astra, Anthropic’s Claude Fable 5.1, and Claude Opus 5.5. The table in the model card includes 19 rows across eight categories. Below are the key results.

The biggest gaps are in white-collar tasks. On Harvey’s legal benchmark, the result is almost three times higher than GPT-6 Astra’s.

The second strong area is long context. Up to 128,000 tokens, all models are close to the ceiling, but in the 256,000 to 1 million token range, Argon maintains 84.2% accuracy. Competitors drop to 65% to 72%.

In agentic coding, the picture is mixed. In the announcement, Google calls the DeepSWE v1.1 result a record, and Argon also leads on Vibe Code Bench (91.9%). At the same time, on FrontierSWE v2 and Terminal-Bench 4.0, the model ranks last out of four. The leader in terminal tasks is Claude Opus 5.5 with 66.4%. In scientific benchmarks, the model leads on LABBench 2 (88.8%) and RiemannBench (76%), but trails GPT-6 Astra and Opus 5.5 on Terminal-Bench Science.

There is also a nuance around the sources of the numbers. Google produced some of Argon’s results itself, while it pulled competitors’ figures from public leaderboards and system cards. The methodology notes which rows were computed by third-party organizations — Vals AI, Zapier, Surge, and others. Independent measurements are worth tracking separately.

How Argon Is Used Inside Google 

Google devoted most of the announcement to internal use cases. According to the Antigravity team, more than 100,000 Alphabet employees use Argon every month in an internal version of the environment. Over four weeks, the share of code written by agents rose by 35%, while the number of rollbacks after submission fell by 30%.

Google gives the following engineering examples:

Chrome. Argon helped separate core web page processing from a legacy browser functions folder of about 6 million lines within Chromium’s 40 million-line codebase. Refactoring hundreds of files took weeks instead of months. According to Google, this will allow pages to move between tabs, side panels, and floating windows without reloading.

YouTube. Load profile analysis showed that nearly 45% of the server costs for the comments frontend came from redundant emoji rendering. After 12 changes, CPU usage was cut in half.

Bug backlog. From a queue of 134 old bug reports, the model closed 97 outdated or duplicate ones in three hours, attaching confirmations from the commit history.

Google also highlights Argon’s adoption beyond development. At Antigravity, more than half of Google Cloud’s sales team, nearly half of the finance department, and more than a third of operations teams use it. For example, a finance leader, over seven workdays and 15.5 hours of their own time, produced an amount of work equivalent to 335 hours of manual labor.

The figures from these case studies are Google’s own data, and it is not yet possible to verify them independently.

Science, Cybersecurity, and Robotics

Beyond software development, Google is positioning Argon as a tool for researchers and security professionals. Here, the company shares results from several fields.

In fusion energy, the Google team, together with Professor Matt Landreman of the University of Maryland, found a new class of counterexamples to Grad’s 1967 hypothesis about stellarator magnetic fields. In medicine, Gamow Labs used Argon to analyze the genomes of newborns with rare lung diseases and found a chromosomal rearrangement that standard software had been missing for years.

To speed up research, Google is offering 10,000 scientists free access to the top-tier Antigravity plan for six months. Applications are already open, and participants will be onboarded gradually based on selection criteria and available capacity..

In robotics, Argon scored 87% on the ERQA spatial reasoning benchmark — the highest among frontier models. In its developer blog, Google shows how the model controls Boston Dynamics’ quadruped robot Spot, first rehearsing movements in the MuJoCo simulator. In another test, Argon watched a video of a person playing cornhole, wrote the simulation code itself, and Apptronik’s humanoid Apollo landed seven bags in a row.

These examples describe the work of teams with early access and expert support. For the average user, they matter mainly as a reference point: which tasks Google expects the model to deliver the biggest gains on.

Model Security

Google lays out the safeguards that were refined between the two launch phases. There are four of them.

1. Abuse prevention. The model refuses to help with cyberattacks and the creation of chemical, biological, radiological, and nuclear weapons, while still supporting legitimate dual-use scientific research. To detect abuse, Google monitors the model’s internal activations.
2. Resistance to prompt injection. These are attacks where malicious instructions are hidden inside the data the model processes. On the Gray Swan benchmark, Argon leads in resilience to these attacks.
3. Misalignment monitoring. A separate system tracks the model’s chain of reasoning and actions, and halts execution if Argon goes beyond the user’s intent. Google urges the industry to maintain transparency in model reasoning.
4. Environment isolation. Before risky training and evaluation stages, sandboxes are isolated and sealed.

The full security report was published alongside the model card. Separately, Google notes that during training it avoided using the outputs of reasoning monitoring, so as not to teach the model to bypass that monitoring.

Pricing and Availability

Argon is more expensive than the Flash models, but Google is launching it at an introductory price. Until November 30, 2026, the following rates apply:

• standard mode — $2 per 1 million input tokens and $10 per 1 million output tokens, including reasoning tokens
• cached input — $0.10 per 1 million tokens, only implicit caching is supported
• batch mode — $1 and $5
• priority mode — $3.6 and $18.

Starting December 1, 2026, all prices will double: standard mode will cost $4 and $20. There is no model on the API free tier.

For developers, Argon is available in the paid Gemini API via AI Studio and Android Studio, and for companies — in Gemini Enterprise and on Google Cloud. In the Antigravity environment — the desktop app and CLI — the model is available only to Google AI Ultra subscribers and users on eligible enterprise plans. The editorial team previously broke down how the environment itself works in a guide to Antigravity.

Regular users get Argon in the Gemini app and AI Mode in Google Search on the AI Pro and Ultra plans. Alongside the model, three new features appeared in the app.

Antigravity in the macOS app. The agentic programming interface is built directly into the Gemini app for Mac, making it easier to switch between chat and code.

Task mode. A “chat — tasks” toggle in the main window. In task mode, Gemini keeps running even when the user is away from the computer.

More than 40 apps. Gemini connects to Adobe, Shopify, and other services — from ordering groceries to booking tables.

Limits in the Gemini app depend on the plan and the complexity of requests: according to Google’s help page, they refresh every five hours within a weekly quota. For Argon, with its long reasoning, this is worth keeping in mind when working on large-scale tasks.

How to Put Argon to the Test: An Interactive Black Hole

Google emphasizes that, compared with previous Gemini models, Argon has made a noticeable leap in spatial reasoning, math accuracy, and visual programming.

Let’s test that in the Gemini app via Canvas. Instead of a library, we’ll use an object that demands precise physics: a supermassive black hole with an accretion disk. The recognizable reference is Gargantua from the film “Interstellar.” Its visualization is based on calculations by physicist Kip Thorne’s team and the studio Double Negative, described in a 2015 scientific paper.

You don’t need to assemble the scene from models and textures: the image is computed entirely in a shader — a program that traces the path of a light ray through curved space for every pixel. That means image quality depends only on how accurately the model implements the physics.

You’ll need a Google AI Pro or Ultra subscription: Argon isn’t available in the free version of the app. For smooth performance, a dedicated graphics card is recommended.

Open gemini.google.com and start a new chat. In the model switcher to the right of the input field, select Gemini 4 Argon.

Then click “+” to the left of the input field and select Canvas. In this mode, the model writes code in a separate panel and immediately runs it in the preview window.

After that, the chat is ready to use. For complex visual projects, Canvas is more convenient than a regular reply: the code and the output are shown side by side, and edits are applied in the same window.

The technical brief describes the physics, appearance, and controls. The model will figure out the calculations and shaders on its own.

Prompt

Create a cinematic real-time visualization of a supermassive black hole in a single HTML file using raw WebGL2 and full-screen GLSL fragment shaders, no external libraries. Render it by ray marching: integrate photon paths in the Schwarzschild metric so that the background and the accretion disk are gravitationally lensed, including the bright photon ring and the image of the far side of the disk bent above and below the black shadow, like Gargantua in the film Interstellar. Accretion disk: thin, from the innermost stable circular orbit out to 12 Schwarzschild radii, with turbulent animated procedural noise orbiting at Keplerian speed (inner parts faster), blackbody temperature colors from white-hot inside to deep orange outside, relativistic Doppler beaming (the approaching side brighter and bluer) and gravitational redshift. Background: a dense procedural starfield with subtle color variation and a faint Milky Way band, also lensed. Post-processing in a separate pass: HDR rendering, ACES tone mapping, a soft wide bloom and subtle film grain. Camera: a slow cinematic auto-orbit at about 10 degrees above the disk plane, mouse drag to orbit, scroll to zoom. A minimal translucent UI panel with sliders for camera inclination, disk brightness, Doppler strength and time speed, plus a quality selector that lowers the ray-marching step count on weak GPUs. Target smooth 60 fps.

The prompt asks the model to do several things at once: numerically integrate the equations of motion for light in curved spacetime, and calculate the disk’s color and brightness according to the laws of physics.

Limitations 

Argon also has limitations that are worth considering before you get started.

Price. After November 30, 2026, API pricing will double to $4 and $20 per 1 million tokens. Long chains of reasoning and responses up to 1 million tokens increase the final bill.

Access. In the Gemini app, the model is available only on AI Pro and Ultra, and in Antigravity — only on Ultra. Argon is not available in the API on the free tier.

Mixed coding results. On FrontierSWE v2 and Terminal-Bench 4.0, the model trails GPT-6 Astra and Claude Opus 5.5. Google calculated some of the figures in the table itself.

Hallucinations and latency. The model card explicitly flags possible hallucinations, as well as slowdowns and timeouts.

Knowledge cutoff. March 2026. It’s best to verify fresh data against primary sources.

What Argon Changes for the Gemini Lineup

With Argon’s release, Google has a flagship model again for the first time in several months. Over the summer, most of the workload was handled by the Flash versions: 3.6, 3.7, and 3.8 launched just a few weeks apart, while 3.1 Pro served as the top model in the Gemini app. Now the lineup has split: Flash — for everyday tasks, Argon — for deep work.

Google calls Argon “just the beginning.” In a thread, the company promised to roll out the rest of the Gemini 4 model family, and in the Antigravity blog — future models that will be better suited for everyday work for a broader range of developers. Timelines and names have not been disclosed yet.

Until they launch, Argon is the flagship and most expensive Gemini model. For AI Pro subscribers, the main use case is complex visualizations and long documents in the Gemini app; for developers, it is long-running agentic tasks in the API and Antigravity, where the model leads in Google’s own tests.