Google Gemini 4 Argon scores 91.3% on ARC-AGI-2, nearing...

Google DeepMind's Gemini 4 Argon has posted a 91.3 percent score on ARC-AGI-2, the benchmark widely used to measure general problem-solving ability in AI systems. On the older ARC-AGI-1 test, Argon hit 98.5 percent at a cost of $0.16 per task. Those numbers put Google closer to the front of the AI race than it has been in months, after a stretch where rivals like OpenAI pulled ahead on the toughest reasoning tests.

The bigger story is ARC-AGI-3, the current flagship benchmark, where Argon scored 97.5 percent at a cost of $4,800. That makes Google the second lab, after OpenAI, to cross the 95 percent threshold on a test designed specifically to resist memorization and brute-force training. Only GPT-6 Astra scores higher, at 99.9 percent. Argon also matched the human efficiency bar in a handful of games, needing 62 percent fewer actions than human testers across 25 public challenges.

But the headline number depends on how the model is tested, and that detail traces back to a dispute from July 2026 involving OpenAI's GPT-5.6 Sol.

Why this matters

The headline number is real, but so is the asterisk. A 97.5 percent score on ARC-AGI-3 and a $4,800 price tag only hold up under the Provider Adapter Harness, a setup that leans on Google's own vendor features. Swap to older benchmarks and the picture gets more useful: 91.3 percent on ARC-AGI-2 at 32 cents a task, 98.5 percent on ARC-AGI-1 at 16 cents, priced off Gemini's $2/$10 per million token rates.

For developers and founders actually budgeting inference costs, those per-task numbers matter more than the headline percentage. Epoch AI's finding that Argon still trails rivals in math and software engineering is the detail worth sitting with. A model can look like it's closed the gap on reasoning benchmarks while still lagging on the tasks that show up in daily engineering work.

Before treating Argon as proof Google has caught the frontier, check which harness produced the number you're looking at, and ask whether coding and math performance, not abstract pattern-matching, is the thing you actually need.