Google opens Gemini 4 Argon to developers, generating 1 million tokens in a single response

Google DeepMind has opened Gemini 4 Argon, the newest member of its frontier reasoning family, to developers. The model was initially made available only to a limited number of experts because of its cybersecurity capabilities; as of mid-October, it has been opened to general developer use via Google Antigravity, AI Studio, and Vertex AI.

What sets Argon apart from other chat models is that it is designed for hours-long autonomous software and 3D generation tasks rather than short question-and-answer exchanges. Google positions the model not as a reactive language model, but as a multimodal decision and generation engine built for long-horizon autonomous engineering.

1 million output tokens in a single response

Until now, when people said “1 million tokens” in large language models, they meant the input window the model could read; the generated response usually topped out around 32,000-64,000 tokens. Gemini 4 Argon can generate 1 million output tokens in a single trajectory. This lets the model complete tens of thousands of lines of procedural geometry, shader math, and architectural structure as a single piece in coding tasks lasting 4 to 11 hours without losing context.

It renders and critiques its own output

The closed-loop self-critique cycle inside Antigravity uses the model’s multimodal vision capability. Argon first writes Three.js code, a WebGL shader, or a Blender Python script; it then examines rendered views of the scene it generated from different camera angles, spots the shortcomings itself, and rewrites the code. This loop can repeat dozens of times without human intervention.

Benchmark scores and pricing

According to the metrics compiled by the source, the model ranks first with 77.9% on the autonomous software engineering test DeepSWE v1.1. It scored 68.9% on the Vals Enterprise Index, which measures enterprise legal, finance, and coding tasks, and achieved a record 68.0% on the autonomous vulnerability repair test CWE-bench v1. Pricing is also notable: 1 million input tokens cost $2, and 1 million output tokens cost $10. In the source article’s comparison, rival frontier models’ output pricing is shown at around $50.

What developers built with Argon

Among the shared examples is a STRIX HX-9 combat helicopter simulation generated entirely from Three.js code, without using off-the-shelf 3D assets or a game engine. Another developer showed an official demo video to the model inside Antigravity, reverse-engineered the workflow, and autonomously modeled temple architecture in Blender consisting of courtyards, tiled roofs, and stone lanterns. A project that built the Eiffel Tower and the surrounding Parisian urban fabric from scratch was completed with 3 hours of autonomous work at an API cost of $21. One of the most striking examples is a photorealistic library scene built with pure mathematical equations inside a single browser file, without using any texture or model files.

Google is also using the model internally: it leverages Argon to autonomously migrate legacy C/C++ folders in the Chromium browser core to Rust, optimize quantum chemistry libraries, and convert raw videos into physically accurate humanoid robot simulations.

Source: WebTuring