The competition in AI is no longer focused on having the largest model.

      TL;DR The prevailing notion that the largest AI model is the best is diminishing, as companies now select models based on task requirements, cost, and control rather than their ranking on leaderboards. Contributing factors include substantial monthly model expenses, the growth of model routing, and the emergence of specialized task-specific agents, which Gartner forecasts will be present in 40% of enterprise applications by the end of 2026, up from below 5%. As capabilities become standardized, the focus shifts to who can provide the least expensive inference.

      For years, the industry operated under the assumption that larger models prevailed. However, this belief is now waning, according to CNBC. Companies are selecting models based on specific tasks, pricing, and control instead of purely their performance metrics. Leaderboards still hold importance, but they are no longer the sole consideration.

      The rationale behind this shift is straightforward. At an enterprise level, model costs can reach millions of dollars monthly.

      The shift towards "good enough" models

      The guiding principle is now to find the most affordable model that meets quality standards. Buyers have realized that many tasks do not require cutting-edge systems. Model routing has developed to streamline this decision-making process, directing requests to the most suitable model. For instance, a summarization task and a complex reasoning task will now be handled by different models.

      Specialized, industry-specific models are beginning to fill in the gaps. Gartner anticipates that by the end of 2026, 40% of enterprise applications will feature task-specific AI agents, a significant increase from less than 5% the previous year.

      The impact of rising costs

      The financial model is no longer sustainable. Despite a drop in per-token costs, enterprise AI expenses have tripled because agentic tools often require significantly more tokens for each task. Buyers are becoming aware of this trend. Nikesh Arora, CEO of Palo Alto Networks, noted that token prices would need to decline by as much as 90% for broader adoption to occur. Some companies have started rationing their use, leading to a trend of "token-minimizing," with organizations imposing strict limits on employee AI spending.

      Where value is headed

      As capabilities become standardized, the profit margin shifts to those who offer the lowest-cost operation. Inference optimization has quietly emerged as one of the most valuable aspects of AI infrastructure. Open and affordable models accentuate this point. Chinese models are approaching the performance of US frontier labs at a fraction of the cost, which sets a ceiling on pricing for merely adequate outputs.

      This situation poses challenges for the scaling hypothesis. Hundreds of billions in capital expenditures were based on the belief that larger models would consistently outperform smaller ones, but buyers are now demonstrating otherwise.

      This does not imply that frontier models are obsolete; rather, the industry is realizing that most tasks are mundane, and such work does not require the most expensive tools available.