Google Deepmind’s new model, Gemini 4 Argon, scored 97.5 percent on the ARC-AGI-3 benchmark using a specific testing setup that costs $4,800 to run.
This result makes Google the second major AI lab to cross the 95 percent mark on a test widely considered the toughest for general problem-solving. Only OpenAI’s GPT-6 Astra scores higher, at 99.9 percent.
The score relies on the “Provider Adapter Harness,” a configuration that uses vendor-specific features like retained reasoning and context compaction. Under this method, Argon needed 62 percent fewer actions than the human baseline across the 25 public games included in the test.
The model matched or beat the speedrun record in three games and came within 10 percent of the best time in six others. On older versions of the benchmark, Argon scored 91.3 percent on ARC-AGI-2 and 98.5 percent on ARC-AGI-1.
Testing conditions drive the numbers
When the model is tested in a “Standard Harness” that only allows simple notes between rounds, the score drops to 49.8 percent at low reasoning levels. This version of the test costs $5,800.
The gap traces back to a dispute in July 2026 when OpenAI’s GPT-5.6 Sol beat Claude Opus 5 on the same benchmark. GPT-5.6 Sol only managed 7.8 percent in the standard environment because it discarded internal reasoning after each action and used an older interface.
Claude was tested under different conditions than GPT-5.6. When tested with OpenAI’s own Responses API, including retained reasoning and smart context compaction, the model jumped to 38.3 percent and beat Opus 5 at 30.2 percent.
OpenAI argued that benchmarks always measure the technical environment too. ARC Prize co-founder Francois Chollet then allowed general, publicly available API settings for the final results.
API problems also prevented Argon from completing all 55 games in the Standard Harness. At low reasoning, the system finished 42 games. At medium reasoning, it completed 46 games. At the highest level, it only completed 19 games, so ARC Prize did not publish a score for that tier. The reported results still cover all 55 games by including partial points from aborted runs.
Progress in general skills, but math and code remain behind
Research institute Epoch AI tracked Google’s progress separately. Argon scored 165 points on the Epoch Capabilities Index, tying Fable 5.1 and landing just behind Opus 5.5 with 167 points and GPT-6 Astra with 166 points.
The index combines over 50 benchmarks into a single capability scale. Argon jumps well ahead of Gemini 3.8 Flash with 157 points and brings Google closer to the top than at any point since Gemini 3.1 Pro in February.
Performance varies by domain. Argon leads several knowledge, agent, and vision benchmarks. Early estimates put it in eighth place on the Math ECI and sixth on the Software Engineering Index.
What it means
Developers and researchers see a model that is significantly more capable in general reasoning than previous versions. However, the reliance on specific API features like retained reasoning means the reported scores may not reflect how the model performs in standard, unassisted environments.
