Gemini 4 Argon: benchmark analysis, spatial reasoning, and real-world coding limitations

For a long time, releases of flagship artificial intelligence models followed a predictable marketing script: developers compiled a test table tailored to their strengths, demonstrated a 2–3% advantage over a competitor, and announced a new technological milestone. With the release of Gemini 4 Argon Google attempted to break this pattern. The model was not created as yet another scaled-down branch like the Flash line, but as a foundational heavyweight system aimed at autonomous scientific and engineering tasks. However, a detailed analysis of independent benchmarks paints a contrasting picture: convincing leadership in some areas is paired with unexpected vulnerability in others.

Independent rankings: what the actual balance of power looks like

Relying exclusively on internal reports from the creators of neural networks has long been considered bad form in the industry. Far more revealing are independent aggregators with open testing methodologies:

The composite index Artificial Analysis Intelligence Index. Unlike narrow leaderboards, it averages the results of dozens of metrics. In this ranking, Argon remains in the upper tier, but its lead over its pursuers looks considerably more modest. This is because some tests in the index are updated more slowly than the architecture of multimodal transformers evolves.

Spatial reasoning: the key to physical intelligence

The model’s most impressive technical achievement was its result on Blueprint Bench 2, developed by the research laboratory Und Labs. In this test, models are fed disparate photographs of the interiors of real apartments, taken from different angles and in difficult lighting conditions. The algorithm’s task is to reconstruct an accurate floor plan of the space, including geometric proportions, the placement of doorways and windows, and the dimensions of furniture.

Gemini 4 Argon outperformed recognized leaders in visual analysis—Claude Opus and GPT-6 Astra. For Google DeepMind, this direction has critical practical significance. Spatial modeling is directly linked to physical intelligence and robotics projects (the RT-X family and Project Astra assistant robots). A robot must do more than generate polite dialogue: it must accurately calculate movement vectors in three-dimensional space, understand the geometry of obstacles, and grasp the physical relationships between objects.

Hallucinations and the 1-million-token output phenomenon

Historically, the Gemini line suffered from a “helpfulness syndrome”: the model confidently fabricated nonexistent quotations, lost track of long-session history, and distorted factual data in specialized topics. In Argon, Google engineers reworked the fact-verification algorithms. In a specialized benchmark for detecting factual errors, the model showed a hallucination rate of just 15%. By comparison, for most competing first-tier systems, this figure fluctuates around 22–30%.

The second architectural surprise is the stated generation limit of up to 1 million tokens in a single response. It is important to distinguish between the context window for receiving information and the output volume:

• The context window determines how much incoming material (books, codebases, hours-long video recordings) the model can retain in memory;
• The generation limit determines the size of the monolithic response the model can produce without being forcibly stopped by the token limit.

If this limit proves resistant to looping and logical degradation in practice, engineers will be able to run end-to-end generations of gigantic software systems, complete monographs, or detailed simulations in a single request, without splitting the work into chains of fragile prompts.

Programming failure: 8th place and Bloomberg skepticism

The model’s only—but extremely painful—trade-off was software development. In Code Arena WebDev the model slipped to eighth place. It trails not only Claude Sonnet and the specialized Codex, but also models from six months ago.

The problem is deep-rooted. When generating autonomous web applications, Argon often makes logical errors in the structure of asynchronous calls, incorrectly links backend modules to the frontend, and struggles to debug other people’s code. For developers planning to use the model in recursive self-improvement cycles (self-improving code loops), 8th place is a serious limiting factor.

Journalists from Bloomberg also confirmed this practical skepticism. Following internal tests in real editorial and analytical scenarios, the publication’s staff noted a curious anomaly: in isolated synthetic tests, Argon achieves outstanding scores, but when faced with chaotic work instructions, incomplete data, and the need for continuous contextual fact-checking, the model quickly loses its advantage and performs no better than the industry’s proven workhorses.