Gemini 4 Argon scores 97.5% on ARC-AGI-3, with a caveat

The score depends on the harness

The 97.5% result came from Provider Adapter Harness, which uses provider-specific features including retained reasoning and context compaction. ARC Prize says Google is the second lab after OpenAI to pass 95% on ARC-AGI-3, a benchmark considered one of the hardest tests of general problem-solving. GPT-6 Astra scored 99.9%.

Across the 25 public games of ARC-AGI-3, Gemini 4 Argon (yellow) and GPT-6 Astra (blue) both needed 62 percent fewer actions than the human baseline (purple) and often came close to the speedrun record (red). | Image: ARC Prize

Source: the-decoder.com

Like Astra, Argon used 62% fewer actions than a human across 25 open games. It matched or beat the speedrun record in three games, and finished within 10% of the best time in another six.

Argon also scored 91.3% on ARC-AGI-2, at $0.32 per task, and 98.5% on ARC-AGI-1, at $0.16 per task. Those costs use the model’s starting prices: $2 per million input tokens and $10 per million output tokens.

Across the 25 public games in ARC-AGI-3, Gemini 4 Argon (yellow) and GPT-6 Astra (blue) both required 62 percent fewer actions than the human baseline (purple) and often came close to the speedrun record (red). | Image: ARC Prize

Source: the-decoder.com

Switch to the Standard Harness, where only simple notes can be carried between rounds, and Argon’s score falls to 49.8% at low reasoning effort. That test cost $5,800.

The gap reflects a broader dispute over what ARC-AGI-3 measures when API features differ. In July 2026, GPT-5.6 Sol beat Claude Opus 5 using two API configurations. In the standard environment, Sol scored just 7.8%: it lost its internal reasoning after each action, and the test used an older interface without Claude’s newer API features. The two models were not tested under equivalent conditions.

Using OpenAI’s Responses API, with retained reasoning and Compaction, GPT-5.6 Sol reached 38.3%, ahead of Opus 5’s 30.2%. OpenAI argued that benchmarks inevitably test the technical environment too. ARC Prize co-founder François Chollet then allowed shared API settings available to all.

The Standard Harness also failed to let Argon complete all 55 games. At low reasoning effort, it finished 42 and scored 49.8%; at medium effort, it finished 46 and scored 46.3%. At maximum effort, it completed only 19, so ARC Prize did not publish a score. The reported figures across the test include partial credit for interrupted attempts.

A broader lead, with clear gaps

Epoch AI’s Epoch Capabilities Index gives Argon 165 points, level with Fable 5.1. Opus 5.5 scored 167 and GPT-6 Astra 166. The index combines more than 50 benchmarks into one capabilities score, much as Artificial Analysis does with its main metric.

Argon is well ahead of Gemini 3.8 Flash, at 157 points. It also brings Google closer to the leaders than any company model since Gemini 3.1 Pro in February, a direction previously noted by Artificial Analysis and Vals.ai.

With an ECI score of 165 points, Google is closer to the current AI frontier led by GPT-6 Astra (166) and Opus 5.5 (167) than it has been since Gemini 3.1 Pro in February 2026. | Image: Epoch AI

Source: the-decoder.com

Epoch AI’s estimates put Argon ahead on several benchmarks for knowledge, AI agents and vision. But the preliminary rankings are weaker in two areas: eighth on Math ECI and sixth on the software engineering index, SWE ECI.

I think the result is strongest as evidence of Google’s progress, not as a clean ranking of model capability. The announcement leaves a more practical question open: how much of Argon’s advantage survives when users cannot rely on provider-specific reasoning and context features? Until that is clear, the 97.5% is as much a score for the testing setup as for the model.