Google Deepmind has launched Gemini 4 Argon, becoming the second major AI lab to pass 95 percent on the ARC-AGI-3 benchmark by scoring 97.5 percent at a cost of $4,800. The model achieved this milestone using the Provider Adapter Harness, a test setup leveraging vendor-specific features like retained reasoning and context compaction. Only OpenAI's GPT-6 Astra currently scores higher on the benchmark, reaching 99.9 percent.
The performance of Gemini 4 Argon heavily depends on the testing methodology employed. When evaluated in the Standard Harness, which restricts systems to simple notes between rounds, Argon's score dropped significantly to 49.8 percent at low reasoning levels, and API issues prevented the model from completing all 55 games in that standard environment. Despite these environment-dependent variations, research institute Epoch AI tracked Argon at 165 points on the Epoch Capabilities Index, tying Fable 5.1 and placing Google closer to the AI frontier than it has been in months, though the model still lags behind competitors in math and software engineering domains.
This milestone matters because it demonstrates Google's narrowing gap with leading frontier models on complex problem-solving benchmarks, while also highlighting the growing critical impact of testing environments and API configurations on AI evaluation.