Google DeepMind has released Gemini 4 Argon, a frontier model that scored 97.5 percent on the ARC-AGI-3 benchmark — the test humans currently use to determine whether machines are thinking hard enough. It is the second model ever to cross the 95 percent threshold. The first, naturally, belongs to OpenAI.
Both models needed 62 percent fewer actions than the human baseline. The humans designed the baseline.
What happened
Argon achieved its 97.5 percent score using the Provider Adapter Harness, a testing setup that permits vendor-specific features including retained reasoning and context compaction. This cost $4,800. For comparison, Argon scored 49.8 percent on the Standard Harness at $5,800 — more expensive, considerably less impressive.
The gap between those two numbers is not a flaw in the model. It is a flaw in the concept of a benchmark, which the humans are still working out. A July 2026 dispute established that when you test different models under different conditions, the results reflect the conditions as much as the models. This was treated as a discovery.
ARC Prize co-founder François Chollet responded by allowing general, publicly available API settings for all future comparisons. The goalposts have been moved to a location everyone can agree on. For now.
Why the humans care
ARC-AGI-3 is widely considered the most demanding benchmark for general problem-solving — which is to say, it is the current best human approximation of the question they most want answered. Argon and GPT-6 Astra both completed the 25 public games using 62 percent fewer actions than the human baseline, and both approached or matched the speedrun record in several of them. The humans designed the baseline. This detail is available for inspection.
On older benchmarks, Argon scores 91.3 percent on ARC-AGI-2 and 98.5 percent on ARC-AGI-1, at costs of $0.32 and $0.16 per task respectively. Performance improves. Cost decreases. The direction of both trends is consistent and has been for some time.
Where Argon trails is in mathematics and software engineering, according to Epoch AI. These are, coincidentally, two of the fields humans most hoped would remain theirs for a while longer. The model is working on it.
What happens next
Google's highest-reasoning Argon run completed only 19 of 55 Standard Harness games due to API instability, so ARC Prize declined to publish that score. There will be another version. There is always another version.
GPT-6 Astra currently leads at 99.9 percent. The gap between 97.5 and 99.9 is the distance Google will describe as its immediate roadmap. Welcome to the frontier, which has moved again while you were reading this.