Gemini 4 Argon scores 97.5% in provider-specific ARC-AGI-3 test

According to an evaluation published on 16 October 2026, Google DeepMind’s Gemini 4 Argon scored 97.5% on the ARC-AGI-3 general problem-solving test using a provider-specific setup. The Decoder reports that this test cost $4,800, citing data from ARC Prize. According to ARC Prize, Google DeepMind became the second AI lab to surpass the 95% threshold on ARC-AGI-3. However, scores from different test setups show that the result depends on the conditions used to measure it.

How did the test setup change the result?

The Provider Adapter Harness can take advantage of certain provider-specific features, including protected reasoning and context-compression capabilities. The more limited Standard Harness allows only simple notes between turns. In this setup, Argon scored 49.8% at the low reasoning level; the test was reported to have cost $5,800.

API issues during Standard Harness runs also affected the number of games completed. At the low reasoning level, 42 of 55 games were completed, compared with 46 at the medium level. At the highest level, only 19 games could be completed, so no score was published for that condition. The reported results also included partial scores from incomplete runs. For this reason, the test setup and the number of completed tasks should be considered alongside the 97.5% result.

What do the other benchmark results show?

Across ARC-AGI-3’s 25 open games, Argon required 62% fewer moves than the human baseline. The model matched or beat the speed record in three games and came within 10% of the best time in six. In earlier test versions, its ARC-AGI-2 score was reported as 91.3%, and its ARC-AGI-1 score as 98.5%. The reported cost per task for those tests was $0.32 and $0.16, respectively.

Argon scored 165 on Epoch AI’s Capabilities Index, tying with Fable 5.1. Opus 5.5 scored 167, GPT-6 Astra 166, and Gemini 3.8 Flash 157. Results by domain paint a different picture from the overall index: in Epoch AI’s preliminary estimates, Argon ranks eighth on Math ECI and sixth on the Software Engineering Index.

What this means for architecture and visualization workflows

These benchmarks do not measure Argon’s ability to handle architectural modeling, image generation, or integration with project software. An ARC-AGI-3 score therefore does not mean that an architectural task will be completed faster or that visualization quality will improve. For architecture firms and visualization teams, the results are meaningful only when tested against their own tasks.

On the cost side, the source reports expenses of $4,800 and $5,800 for the ARC-AGI-3 test setups; these are not presented as direct product usage rates. It also lists starting API prices of $2 per million input tokens and $10 per million output tokens. The source provides no information about hardware requirements, licensing scope, or compatibility with architectural software. Until these points are clarified, there is not enough information to make a decision about existing production workflows.