We tested frontier models on slide creation (slidely.ai)

A recent evaluation of AI slide creation models revealed that OpenAI's GPT-6 Astra significantly outperformed competitors in both quality and cost-effectiveness, making it the top choice in the Slidely harness. Astra achieved an Arena score of 1090.5, surpassing Fable 5.1 by 40 points, and is less than half the cost of Fable, offering an ideal combination of high-quality output at $1.16 per slide. This marks a notable advancement over previous generations like GPT-5.6, which have historically lagged behind Google's and Anthropic’s offerings. Astra’s exceptional spatial reasoning capabilities allow for better content adaptation while maintaining the layout's integrity, improving text contrast, and efficiently utilizing space, especially in high-density slides. The methodology for evaluating the models was rigorous, involving blind pairwise comparisons among three AI reviewers, ensuring that results correlated well with human expert evaluations. Notably, Astra produced the fewest output tokens, suggesting efficiency in generating high-quality slides. As the AI/ML community looks to refine models for more complex tasks, Astra’s performance indicates a significant leap in generative capabilities, paving the way for future innovations and benchmarks in AI-driven creative functions. This study also highlights the need for more challenging benchmarks to further advance model evaluations in slide creation.