Google DeepMind's new flagship model, Gemini 4 Argon, has become only the second AI system to score above 95% on ARC-AGI-3, a benchmark widely seen as the hardest test of general problem-solving. It hit 97.5%, behind only OpenAI's GPT-6 Astra at 99.9%. On the face of it, that puts Google right back at the front of the AI race.
Look a little closer and the picture gets more interesting.
The test setup changes everything
That 97.5% was achieved using what the ARC Prize team calls the "Provider Adapter Harness", a testing setup that lets each company use its own special features, such as keeping the model's earlier reasoning in memory and compressing long conversations. It cost about $4,800 to run.
In the "Standard Harness", which only allows the model to keep simple notes between rounds, Argon scored 49.8% at its lowest reasoning setting. API problems also stopped it finishing every game in the standard test, so some of those results include partial marks from runs that broke off.
The two-track system grew out of a row in July, when OpenAI argued that the standard setup unfairly threw away its model's reasoning after each move. ARC Prize co-founder François Chollet then allowed companies to use general, publicly available settings. Both numbers are "real". They just measure different things.
Strong, but not best at everything
Independent researchers at Epoch AI, who combine more than 50 benchmarks into one score, put Argon at 165 points. That ties it with one rival and places it just behind Anthropic's Opus 5.5 (167) and GPT-6 Astra (166). It is the closest Google has been to the top since February. But Epoch's early estimates place Argon only eighth for maths and sixth for software engineering, even as it leads on several knowledge, agent and vision tests.
How to read the next AI headline
Three simple questions cut through most benchmark hype. First, which version of the test was used, and could rivals use the same setup? Second, what did it cost? A score bought with thousands of dollars of computing is not the same as one you will get on your phone. Third, what is the model actually good at? A single headline number hides big differences between maths, coding and everyday tasks.
Gemini 4 Argon is clearly a serious model. But the most useful lesson from its launch is that in AI, the small print matters as much as the score.