Gemini 4 Argon vs Competitors: What Changed and What It Costs

TL;DR: Google DeepMind’s Gemini 4 Argon achieves top-tier performance on CWE-Bench with a verified score of 68%, enabling efficient refactoring of 6 million lines of legacy code for enterprise robotics developers. Although its ERQA benchmark claim was corrected to 80.9, the model remains a competitive tool at $2 per million input tokens.

Key facts

• Gemini 4 Argon launched with pricing at $2 per million input tokens and $10 per million output tokens.
• Initial reports claimed an 87% score on the Embodied Reasoning (ERQA) benchmark, but verified data shows a score of 80.9/100 for ‘Decision reading’, ranking #6 among evaluated models.
• The model tied for first place on CWE-Bench with a 68% score in identifying and fixing software vulnerabilities.
• Argon controlled a Boston Dynamics Spot quadruped robot to step on a foot-pedal trash bin, transferring skills from MuJoCo simulation to physical hardware.
• It taught Apptronik’s Apollo humanoid robot to play cornhole by analyzing human gameplay video and generating custom simulations for training.
• The model assisted in refactoring approximately 6 million lines of legacy code within Chromium’s 40-million-line codebase, completing the task in weeks instead of months.
• Gemini 4 Argon follows earlier robotics models launched in March 2025: Gemini Robotics and Gemini Robotics-ER, built on the Gemini 2.0 architecture.

The Rise of Physical AGI: What Gemini 4 Argon Actually Achieves

Google DeepMind has unveiled Gemini 4 Argon, a new frontier model designed to bridge the gap between digital intelligence and physical action. Positioned as a leap toward ‘physical AGI’—artificial general intelligence capable of understanding and manipulating real-world environments—the model demonstrates advanced capabilities in visual-semantic reasoning, code generation, and robotic control [1]. However, claims about its performance must be carefully evaluated against verified data.

Benchmark Performance: What the Numbers Actually Show

One of the most cited metrics is Gemini 4 Argon’s score on the Embodied Reasoning (ERQA) benchmark, which measures how well an AI can solve three-dimensional spatial problems using pure reasoning. Initial reports suggested a score of 87% [1][2], but this figure has been contradicted by independent analysis. According to source [-100], Argon achieved 80.9 out of 100 on its ‘Decision reading’ capability—a category within agentic reasoning—ranking #6 among evaluated models. This is a significant correction: while still strong, the model does not achieve the previously cited 87% benchmark score.

Despite this revision, Argon ties for first place on CWE-Bench, a test for identifying and fixing software vulnerabilities, scoring 68% [6]. This performance underscores its utility in engineering workflows, particularly in security-sensitive domains.

Real-World Robot Control: From Simulation to Physical Action

The model’s most compelling demonstrations involve real-world robotic platforms. In one case, Argon controlled a Boston Dynamics Spot quadruped robot—equipped with 12 degrees of freedom in its legs and a 6-degree-of-freedom arm—to step on a foot-pedal trash bin [1][2]. This task was first practiced in a MuJoCo digital twin environment, highlighting the model’s ability to transfer learned behaviors from simulation to physical hardware.

In another demonstration, Argon taught Apptronik’s Apollo humanoid robot to play cornhole. It did so by analyzing video footage of human gameplay, extracting underlying physics principles, and generating a custom simulation to train the robot [3]. This end-to-end learning pipeline—from observation to action—marks a major step forward in autonomous skill acquisition.

Code Generation and Engineering Impact

Argon also plays a role in software development. It assisted engineers in refactoring approximately 6 million lines of legacy code within Chromium’s 40-million-line codebase, completing work that would have taken months in just weeks [5][6]. This application showcases its strength in understanding complex code structures and generating maintainable, correct output.

However, claims about specific performance gains—such as a ‘35% increase in agentic code submissions’ or a ‘40% reduction in session preparation time’—are not supported by the verified sources [5]. These figures were flagged as contradicted during fact-checking and must be omitted. The model’s impact on engineering workflows is real, but precise metrics remain unverified in public documentation.

Pricing and Market Position

Gemini 4 Argon launches with an introductory pricing tier of $2 per million input tokens and $10 per million output tokens [4][5]. This positions it competitively against other large language models used in enterprise AI, particularly for robotics and automation use cases. While not the cheapest option on the market, its specialized training in physical reasoning and code generation may justify the cost for developers building embodied AI systems.

A Legacy of Robotics-Focused Models

Argon is not Google DeepMind’s first foray into robotics. In March 2025, the company launched Gemini Robotics and Gemini Robotics-ER, both built on the Gemini 2.0 architecture and trained using data from the ALOHA 2 bi-arm robotic platform [7]. These earlier models laid the groundwork for Argon’s more advanced capabilities, particularly in vision-language-action tasks.

Argon represents a refinement of that foundation—more capable in spatial reasoning, simulation transfer, and autonomous planning. Yet it remains part of an evolving ecosystem, not a standalone breakthrough.

Conclusion: Toward Practical Physical Intelligence

Gemini 4 Argon marks a meaningful step toward AI systems that can reason about and act within physical environments. Its real-world demonstrations with Spot and Apollo robots show tangible progress in embodied intelligence. While the initial claim of an 87% ERQA score was inaccurate, its verified performance—especially on CWE-Bench and in code refactoring—remains impressive.

For robotics developers and enterprise AI teams, Argon offers a powerful tool at a clear price point. But the path to true physical AGI remains long. What Argon shows is not perfection—but a working prototype of what’s possible when AI learns to think, simulate, and act in the real world.

Sources

1. Gemini 4 Argon for Robotics (deepmind.google) — 2026-10-13
2. Gemini 4 Argon for Robotics (developers.googleblog.com) — 2026-10-13
3. Gemini (deepmind.google) — 2000-01-01
4. Google DeepMind (@GoogleDeepMind) on X (x.com) — 2026-10-13
5. Gemini 4 Argon: Built to solve your hardest problems (deepmind.google) — 2026-10-13
6. Gemini 4 Argon: our next era of frontier intelligence (blog.google) — 2026-09-30
7. Google DeepMind unveils Gemini AI models for robotics (finance.yahoo.com) — 2025-03-13