Can tech companies learn to love cheaper AI models?

The rapid evolution of artificial intelligence has long been predicated on the notion that larger models yield superior performance, with the most advanced models dominating the market. However, this assumption is being challenged as rising costs prompt organizations to reassess their reliance on expensive models. There is a growing trend towards considering smaller, more affordable alternatives, and the implications of this shift could be profound. Notable figures in the tech industry, such as Brian Armstrong, co-founder of Coinbase, predict that a significant majority of tasks will transition to these cheaper models within the next year to year and a half. Armstrong notes that while the demand for AI capabilities remains virtually limitless, up to 80% of workloads could soon be handled by models that are up to 99% less expensive. In contrast, only 20% of tasks may still require the latest and most powerful models for high-stakes applications. If Armstrong's forecast materializes, it could represent a monumental transformation in the AI landscape. Historically, companies have focused on the quality of their AI solutions, often defaulting to the most sophisticated models available. However, if less costly models can effectively perform the same tasks without compromising quality, it would disrupt the current economic framework of the industry, particularly impacting major players like OpenAI and Anthropic as they approach their initial public offerings. The question looming over this potential shift is whether companies are prepared to embrace smaller models. Initial experiments hint that, when optimized correctly, these lower-cost models can deliver comparable quality. A recent case study involving the legal AI platform Harvey demonstrated that it was possible to reduce inference costs by three times without sacrificing output quality. Collaborating with Fireworks AI, Harvey successfully combined advanced models to achieve significant cost savings while maintaining performance. Harvey's co-founder, Gabe Pereyra, emphasized that while quality remains paramount—especially in legal applications—the definition of quality is evolving. It is now more about identifying the most efficient model that yields accurate results rather than simply opting for the most powerful option. This trend is often oversimplified as a battle between major labs and open-source models, but the true dichotomy lies between large and small models. Transitioning from a high-capacity model like GPT-5.5 to a smaller alternative can yield cost savings without a drop in performance. Currently, there is an intense price competition between proprietary models from large labs and independently offered open-source solutions. Although it may seem intuitive to choose the most resource-efficient option, this perspective contradicts the scaling-centric mindset that has prevailed in the industry. Companies have aggressively pursued the development of the most compute-intensive models, fueled by substantial investments, leaving little incentive for clients to consider alternatives. As token prices rise and investor support wanes, users are now confronted with the financial implications of their choices. Whether this newfound cost sensitivity will lead enterprises to adopt smaller models remains uncertain. Companies may opt to reduce usage, limit context, or forgo less promising projects altogether instead. However, if it is proven that smaller models can perform just as effectively, the burgeoning demand for inference could be significantly tempered, prompting a reevaluation of the justification for the costs associated with training cutting-edge models.