Google Gemini 4 Argon: 97.5% Benchmark Score and the Testing Gap

What happened

Google DeepMind's Gemini 4 Argon scored 97.5 per cent on the ARC-AGI-3 benchmark, making Google the second AI lab after OpenAI to pass the 95 per cent mark on that test. The result places Argon closer to the current frontier than any previous Google model, according to reporting by The Decoder.

The headline score depends heavily on how the model is tested. Under the standard evaluation harness, Argon's score falls to 49.8 per cent. The same reporting indicates that Argon also trails rival models in mathematics and software engineering, so its strengths are uneven across task types rather than uniformly high.

The development is a model-level update rather than a deployed product, so its practical effect will depend on how Google makes Argon available and how organisations choose to use it.

Why it matters

For organisations adopting AI, the gap between the headline score and the standard-harness result is the most useful signal in this story. A single benchmark figure reflects the conditions under which it was produced, and those conditions can change the answer substantially. Leaders evaluating models for customer-facing or internal services should treat a headline number as a starting point, not a verdict on suitability.

The uneven performance across mathematics and software engineering also matters for digital transformation programmes. Teams building automated workflows, service tools or internal engineering support will need to match models to specific tasks, and to test them against the processes they actually run.

By the numbers

• 97.5%: Gemini 4 Argon's score on ARC-AGI-3 under the headline test setup.
• 49.8%: Argon's score under the standard harness, roughly half the headline figure.
• 95%: the threshold that Google crossed as the second lab after OpenAI.

The Renascence take

Most coverage of model launches leads with the top-line number, and most buyers anchor on it. The more useful question is how far that number moves when the conditions change, because that spread tells you how much to trust it.

A benchmark score is a framing device before it is a measurement, and a 48-point swing between test setups is exactly the kind of variance that anchors decisions. The principle underneath is that context shapes perception: the same model can look exceptional or ordinary depending on the frame. A customer-obsessed operator should stop asking which model scores highest and start asking how each model performs on its own high-volume customer journeys, under its own conditions. Track the variance and the failure cases, not only the peak.