The competition in AI is no longer centered around having the largest model.

      The belief that larger AI models are always the best is faltering, as companies are now opting for models based on their task suitability, cost, and control rather than simply their rank on leaderboards. This shift is largely driven by high operational costs, with model bills reaching millions monthly, the emergence of model routing, and the introduction of specialized task-specific agents. Gartner predicts that by the end of 2026, 40% of enterprise applications will incorporate these specific AI agents, up from less than 5% the previous year. As capabilities become more standardized, the focus shifts to who can perform inference at the lowest cost.

      For years, a prevailing notion in the industry was that the largest model would prevail. However, according to CNBC, this belief is now changing. Organizations are evaluating models based on their tasks, expenses, and control rather than their position in benchmarks. While being at the forefront remains relevant, it is no longer the sole criterion for selection.

      The underlying reason for this shift is straightforward: enterprise expenses for models have soared into the millions of dollars monthly. The current trend is to utilize the most affordable model that meets quality standards. Companies have realized that many tasks do not require cutting-edge systems.

      Model routing has been introduced to streamline this decision-making process, directing requests to the most suitable model based on the task. Consequently, a summarization task and a complex reasoning task will be handled by different models. Additionally, specialized, industry-focused models are emerging to fill remaining gaps. Gartner anticipates that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, a significant increase from under 5% previously.

      The rising costs compelled this change. Even though per-token prices have fallen, enterprise AI expenses have tripled, primarily because agent-driven tools require significantly more tokens for each task. Buyers have taken note, with Palo Alto Networks' CEO Nikesh Arora indicating that token prices must drop by up to 90% for broader adoption to occur. Some companies have begun to restrict usage, implementing measures to minimize token consumption and limit employee spending on AI.

      As capabilities become commoditized, profit margins will favor those who can provide the service at the lowest cost. Inference optimization has quietly become one of the most critical components of AI infrastructure. The availability of affordable and effective models, including those from China, is intensifying competition, thereby capping prices for standard outputs.

      This development presents challenges to the previously held belief that larger models would continue to outperform smaller ones decisively. Companies have begun to act otherwise. However, this does not signify the end for frontier models; instead, the industry is realizing that many tasks are mundane and do not necessitate the highest-end tools available.