Antigravity Agent Preview

Baseline established for Antigravity Agent Preview

This inaugural benchmark establishes the baseline performance profile for Google Gemini's Antigravity Agent Preview. The model demonstrates strong reasoning capabilities with a score of 78.3 on the GPQA benchmark, indicating solid performance on graduate-level scientific questions. Mathematical reasoning shows competence at 71.2 on MATH-500, while coding ability registers at 74.8 on HumanEval, suggesting reliable performance across technical domains. Instruction following achieves 69.4 on IFEval, showing room for improvement in adhering to complex constraints. The model exhibits a 62.7% win rate on style control tasks, indicating moderate success in matching desired output formats. Multilingual performance is notably weaker at 52.1 on MMMLU, suggesting English-centric optimization. Response time averages 2.8 seconds with relatively low variability, providing consistent user experience. As this is the first evaluation window, these metrics serve as the reference point for tracking future performance trends. Users should expect strong technical reasoning with some limitations in instruction adherence and non-English language tasks.