Google's Gemini 4 Argon Nears AI Frontier with Record ARC-AGI-3 Score
Google DeepMind's Gemini 4 Argon scored 97.5 percent on the ARC-AGI-3 benchmark, becoming the second lab after OpenAI to cross the 95 percent threshold, but its performance drops sharply under standard testing conditions and trails rivals in math and software engineering.
Google DeepMind's latest model, Gemini 4 Argon, has achieved a score of 97.5 percent on the ARC-AGI-3 benchmark, making Google the second artificial intelligence laboratory after OpenAI to surpass the 95 percent mark on that test. The result positions Argon closer to the current AI frontier than any Google model in recent months, according to an analysis of the benchmark data.
The ARC-AGI-3 benchmark is designed to measure abstract reasoning and general intelligence in AI systems, and crossing the 95 percent threshold has become a significant milestone for frontier labs. OpenAI was the first to reach that level, and Google's achievement with Argon suggests the company is narrowing the gap with its main competitor.
However, the strong headline figure comes with a major caveat. When Argon is evaluated using the standard testing harness — the conventional setup used to compare models fairly — its score falls dramatically to just 49.8 percent. That swing of nearly 48 percentage points indicates that the model's performance is highly sensitive to how the test is configured, raising questions about how much the 97.5 percent result reflects genuine reasoning ability versus favorable testing conditions.
Beyond the ARC-AGI-3 benchmark, Argon also trails its rivals in mathematics and software engineering tasks. Those areas have become key battlegrounds for AI developers, as they are seen as indicators of a model's practical usefulness for coding, quantitative analysis, and complex problem-solving. The fact that Argon lags in these domains suggests that Google's model, while impressive on abstract reasoning, may not yet match the best available systems across the board.
The results highlight a broader challenge in evaluating frontier AI models: benchmark scores can vary widely depending on the testing methodology, and a single strong number does not necessarily translate into consistent real-world performance. For researchers and developers comparing models, the gap between Argon's best-case and standard-harness scores is a reminder that evaluation conditions matter as much as raw capability.
Google DeepMind has not yet commented publicly on the detailed breakdown of Argon's performance or on the reasons behind the discrepancy between the two testing setups. The company has been under pressure to demonstrate that its Gemini line can compete with the most advanced models from OpenAI and other leading labs, and the ARC-AGI-3 result offers both encouragement and a note of caution.
For now, Argon's 97.5 percent score on ARC-AGI-3 stands as a notable achievement and a sign that Google remains a serious contender in the race toward more capable AI systems. But the sharp drop under standard conditions, combined with weaker results in math and software engineering, suggests that the path to the frontier is not yet fully secured.
8
