Google Deepmind’s Gemini 4 Argon has moved close to the top tier of AI models in recent evaluations. Its headline result on ARC-AGI-3 is strong, but the large difference between testing setups makes the score difficult to interpret without context.
Argon’s strongest result comes with a specialized setup
Gemini 4 Argon reached 97.5 percent on ARC-AGI-3, at a reported cost of $4,800, when evaluated with the Provider Adapter Harness. That setup permits vendor-specific features, including retained reasoning and context compaction.
ARC Prize says Google is the second AI lab to pass 95 percent on the benchmark, after OpenAI. GPT-6 Astra is the only model listed above Argon, with a score of 99.9 percent. ARC-AGI-3 is described as a demanding test of general problem-solving.
On the 25 public games, Argon used 62 percent fewer actions than the human baseline. It matched or surpassed the speedrun record in three games and came within 10 percent of the best time in six more. Those results measure performance in the games; the harness used to produce them remains part of the story.
The testing method changes the score
Under the Standard Harness, which allows only simple notes between rounds, Argon scored 49.8 percent at low reasoning, with a reported cost of $5,800. That is far below its Provider Adapter Harness result, illustrating how much a benchmark score can depend on the technical conditions around a model.
The article traces the debate to a July 2026 dispute over ARC-AGI-3 testing. GPT-5.6 Sol scored 7.8 percent in the standard environment, which discards internal reasoning after each action and used an older interface without newer Claude API features. When tested through OpenAI’s Responses API with Retained Reasoning and smart context Compaction, GPT-5.6 Sol scored 38.3 percent, ahead of Opus 5 at 30.2 percent.
OpenAI argued that benchmark results also reflect the technical environment. ARC Prize co-founder Francois Chollet then allowed general, publicly available API settings. The episode shows why comparisons require attention to the tools and settings available to each model, not only the final percentage.
Argon also had trouble completing the Standard Harness evaluation. At low reasoning, it finished 42 of the 55 games and scored 49.8 percent; at medium reasoning, it finished 46 and scored 46.3 percent. At the highest reasoning level, it completed only 19 games, so ARC Prize did not publish a score. Reported results across the 55 games include partial points from runs that were aborted.
Broader ratings show progress, with uneven strengths
Epoch AI gave Argon 165 points on its Epoch Capabilities Index, tying Fable 5.1. Opus 5.5 scored 167 and GPT-6 Astra 166. The index combines more than 50 benchmarks into a single capability scale.
Argon’s score puts it well ahead of Gemini 3.8 Flash, which received 157 points, and brings Google closer to the leading models than at any point since Gemini 3.1 Pro in February. Artificial Analysis and Vals.ai had previously confirmed the broader upward trend.
That overall progress does not mean Argon leads in every area. Epoch AI’s early estimates place it among the stronger models on several knowledge, agent, and vision benchmarks, but eighth on the Math ECI and sixth on the Software Engineering Index. The profile is therefore mixed: high performance on some evaluations, with room to catch up in math and software engineering.
What the results can and cannot show
Argon’s ARC-AGI-3 score under the Provider Adapter Harness signals a high level of performance under that setup. Its much lower Standard Harness result, alongside incomplete runs, means the headline number should be read alongside the testing conditions and operational limits.
The older benchmark versions offer additional context: Argon scored 91.3 percent on ARC-AGI-2 and 98.5 percent on ARC-AGI-1. Those figures were calculated using the introductory price of $2 per million input tokens and $10 per million output tokens, with reported costs of $0.32 per task and $0.16 per task respectively. Taken together, the evaluations suggest Google has narrowed the gap to top models, while also showing why no single benchmark captures an AI system’s overall capability.