Google Argon: 97.5% on ARC, but 49.8% on standard test

Google's Gemini 4 Argon model is nearing the top of an aggregated index, achieving a score of 97.5% on the ARC-AGI-3. However, this level relies on a configuration leveraging API functions, while the score drops to 49.8% in a standard framework. Epoch AI credits Argon with 165 points on the ECI, while noting weaknesses in mathematics and software engineering.

ECI: 165 points for Argon, behind Astra and Opus

Epoch AI tracks the progress of Google's models and assigns Argon 165 points on its Epoch Capabilities Index, tying it with Fable 5.1 and just behind GPT-6 Astra at 166 and Opus 5.5 at 167. The ECI aggregates over 50 benchmarks into a single measure, similar to the approach taken by Artificial Analysis for its main score. With 165 points, Google is closer to the current frontier than it has been since Gemini 3.1 Pro in February 2026. Argon also surpasses Gemini 3.8 Flash, which scored 157. According to Epoch AI, Argon ranks first on various benchmarks related to knowledge, agents, and vision, but initial evaluations place it eighth on the ECI Math and sixth on the SWE ECI index.

Standard framework drops Argon to 49.8%

In the Standard Harness, which only allows simple scores between rounds, Argon achieves 49.8% in low reasoning mode for a cost of $5,800. API difficulties prevented the model from completing all 55 games, according to testers. In low reasoning, it finished 42 games (49.8%) and in medium mode, 46 games (46.3%). At the highest level, only 19 games were completed, leading ARC Prize to refrain from publishing a score. Nevertheless, the aggregated results cover 55 games by incorporating partial points from interrupted attempts.

Controversy 2026: API parameters and comparability

In July 2026, OpenAI's GPT-5.6 Sol outperformed Claude Opus 5 on the ARC-AGI-3 thanks to two API parameters. In the standard environment, GPT-5.6 Sol only reached 7.8%, a framework that rejected internal reasoning after each action and used an older interface lacking recent Claude API features, placing the models in different conditions. With OpenAI's API Responses, incorporating Conserved Reasoning and Context Compaction, GPT-5.6 Sol rose to 38.3% and surpassed Opus 5 at 30.2%. OpenAI defended the notion that benchmarks also measure the technical environment, and Francois Chollet subsequently opened the use of general and public API parameters.

97.5% on ARC-AGI-3 with an API-utilizing harness

Gemini 4 Argon displays 97.5% on the ARC-AGI-3 for $4,800 using the Provider Adapter Harness, which mobilizes features like conserved reasoning and context compression. ARC Prize thus positions Google as the second lab, behind OpenAI, to cross the 95% threshold, while GPT-6 Astra stands at 99.9%. Of the 25 public games in the ARC-AGI-3, Argon and Astra required 62% fewer actions than the human benchmark and often approached the speedrun record. Argon matched or beat this record in three games and stayed within 10% of the best time in six others.

Previous ARC versions: high scores and detailed costs

On the ARC-AGI-2, Argon achieved 91.3% with an estimated cost of $0.32 per task. For the ARC-AGI-1, the model shows a score of 98.5% with a cost of $0.16 per task. These values are calculated from an initial rate of $2 for one million tokens in input and $10 for one million tokens in output.