Gemini 4 Argon scores 97.5% on ARC-AGI-3 with vendor features, but only 49.8% in the standard harness

Google DeepMind's Gemini 4 Argon scored 97.5% on ARC-AGI-3 at a cost of $4,800 using the "Provider Adapter Harness," according to ARC Prize. That makes Google the second lab after OpenAI to pass 95% on the benchmark. Only OpenAI's GPT-6 Astra scores higher, at 99.9%. The result depends heavily on the test setup, and Epoch AI data shows Argon behind rivals in math and software engineering.

ARC-AGI results

The Provider Adapter Harness uses vendor-specific features such as retained reasoning and context compaction. Across the 25 public ARC-AGI-3 games, ARC Prize reports that Argon needed 62% fewer actions than the human baseline, the same figure as GPT-6 Astra. Argon matched or beat the speedrun record in three games and came within 10% of the best time in six others.

On earlier versions of the benchmark:

• ARC-AGI-2: 91.3% at $0.32 per task
• ARC-AGI-1: 98.5% at $0.16 per task

Those costs are based on the introductory API price of $2 per million input tokens and $10 per million output tokens. Context window size, parameter count and training cutoff were not disclosed in the source material.

Harness choice changes the result

In the Standard Harness, which allows only simple notes between rounds, Argon drops to 49.8% at low reasoning, at a cost of $5,800. API problems kept Argon from completing all 55 games in this setup, according to the testers:

• Low reasoning: 42 games completed, 49.8%
• Medium reasoning: 46 games completed, 46.3%
• Highest reasoning: 19 games completed, so ARC Prize published no score

The reported scores still cover all 55 games because they include partial points from aborted runs.

The split traces back to a July 2026 dispute. OpenAI's GPT-5.6 Sol scored only 7.8% in the standard environment, which discards internal reasoning after each action and used an older interface that lacked newer Claude API features. Claude Opus 5 was therefore tested under different conditions. With OpenAI's Responses API, including Retained Reasoning and Compaction, GPT-5.6 Sol reached 38.3% and beat Opus 5 at 30.2%. OpenAI argued that benchmarks always measure the technical environment as well. ARC Prize co-founder Francois Chollet then allowed general, publicly available API settings.

Epoch AI: closer to the frontier, with gaps

Epoch AI gives Argon 165 points on the Epoch Capabilities Index (ECI), which combines more than 50 benchmarks into one scale. That ties Fable 5.1 and sits just behind Opus 5.5 (167) and GPT-6 Astra (166). Argon is well ahead of Gemini 3.8 Flash (157). According to Epoch AI, it is Google's closest position to the frontier since Gemini 3.1 Pro in February 2026. The source says Artificial Analysis and Vals.ai had previously shown the same trend.

Performance varies by domain. Epoch AI says Argon leads several knowledge, agent and vision benchmarks, but early estimates place it eighth on the Math ECI and sixth on the Software Engineering Index (SWE ECI).

What this means

The 97.5% headline number is not comparable to scores from other setups. The gap between 97.5% and 49.8% for the same model shows that harness design, and whether a vendor's own API features are allowed, can matter more than the model. Comparisons across labs are only meaningful when the harness is held constant, and the Standard Harness results, hampered by API problems, are incomplete.

The ECI result is the sturdier signal. A 165 score, within two points of the leaders, indicates Google has largely closed the aggregate gap. The weaker math and software engineering rankings matter for builders choosing models for coding agents or quantitative work, where Argon is not yet a leading option. The $2/$10 introductory pricing suggests Google is competing on cost as well, though the source does not say how long the introductory rate lasts.