Google's Gemini 4 Argon is closer to the AI frontier than any Google model in months

Google DeepMind has released Gemini 4 Argon, a model that achieved a 97.5% score on the ARC-AGI-3 benchmark, a test designed to measure abstract reasoning and generalization. This result places Google as the second organization, after OpenAI, to surpass the 95% threshold on this particular evaluation.

However, the headline number comes with a significant caveat. When evaluated under a standard test harness, Argon's score falls to just 49.8%, suggesting that performance is highly sensitive to how the benchmark is configured. This discrepancy raises questions about the robustness of the result and whether the higher score reflects genuine reasoning ability or adaptation to a specific test setup.

ARC-AGI-3 is widely regarded as one of the more challenging benchmarks for measuring general intelligence in AI systems, as it requires models to solve novel problems without prior exposure to similar tasks. Crossing the 95% mark is considered a meaningful milestone in the pursuit of systems that can reason abstractly.

The wide variance in scores depending on harness configuration underscores an ongoing challenge in AI evaluation: results can shift dramatically based on minor changes in testing methodology. Researchers and practitioners will need to carefully consider how benchmarks are administered before drawing conclusions about a model's true capabilities.

For the broader AI industry, Argon's performance suggests that frontier models are continuing to close the gap on abstract reasoning tasks. Whether this translates into practical improvements in real-world applications remains to be seen, but the benchmark result is a notable data point in the ongoing race toward more general-purpose AI systems.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.