Nathan Benaich, founder of Air Street Capital, has been publishing the annual 'State of AI' report for the past seven years, offering valuable insights into the evolution and deployment of AI technology.

The latest report reveals a significant reduction in the cost of using trained AI models, known as 'inference', driven by heightened competition and optimization efforts from companies like OpenAI and Anthropic.

This rapid decrease in costs is largely attributed to competitive pressures focused on performance and pricing.

For instance, OpenAI's GPT-4o now costs 100 times less per token compared to its initial launch in March 2023, while Google's Gemini 1.5 Pro has seen a 76% reduction in cost per output token since its launch in February 2024.

Engineers at OpenAI and Anthropic are optimizing AI models to enhance their efficiency on GPU clusters, contributing to these cost reductions.

While OpenAI's GPT-4o continues to lead in reasoning tasks, the new o1 'Strawberry' model demonstrates a mix of capabilities that vary by task.

Moreover, researchers are successfully developing smaller AI models that can perform comparably to larger counterparts in tasks like dialogue and coding, while being more economical to operate.

These trends are making AI-based solutions more economically viable, potentially improving return on investment for businesses leveraging generative AI.

Despite the competitive landscape, OpenAI retains an edge in reasoning tasks with its o1 'Strawberry' model, which showcases a blend of strong and weak logical abilities depending on the specific task.

However, OpenAI's dominance in the AI sector is waning, as competitors like Anthropic, Google, X, and Meta have achieved similar or superior performance on various benchmarks.

Overall, the competitive dynamics in the AI industry are reshaping the landscape, with significant cost reductions and advancements in model performance.