Tech Giants Confront the Challenge of More Cost-Effective AI Models

A Paradigm Shift in the AI Industry

The rise of artificial intelligence is based on a widely accepted idea: the larger a model is, the more powerful it becomes, and thus, the better it performs. However, this assumption may be called into question as the industry prepares to explore the implications of such a shift.

Rising costs are already prompting users to consider smaller, more affordable models. This movement towards more economical solutions is still recent, and while its effects on the industry remain uncertain, they could be significant.

Brian Armstrong's Vision

Brian Armstrong, co-founder of Coinbase, anticipates a major shift towards less expensive models. According to him, the demand for artificial intelligence is nearly limitless, but he predicts that 80% of tasks will be performed by models that are 99% cheaper in the next 12 to 18 months. The remaining 20% will still require the most advanced models for specific optimizations.

If this prediction comes true, it could radically transform the AI landscape.

Competition on Quality

Historically, AI companies have focused on quality, which has driven them to adopt the most advanced models available. If tasks can be accomplished with less expensive models without compromising quality, it would represent a major upheaval in the AI economy. The savings would primarily come from large labs, which could have financial repercussions for companies like OpenAI and Anthropic, especially as they approach their IPOs.

Promising Tests with Smaller Models

Initial tests indicate that smaller, cheaper models can be used without loss of quality, provided the system is well-organized. For example, Harvey, an AI tool for the legal sector, successfully reduced its inference costs by three times without affecting quality. In collaboration with Fireworks AI, Harvey utilized Claude Opus and GLM 5.1 from Fireworks, reserving Opus for the most demanding tasks. This approach significantly reduced server time and overall costs.

Redefining Quality

Gabe Pereyra, co-founder of Harvey, emphasized that quality remains paramount, especially in the legal field. However, the definition of quality is evolving: it is no longer just about using the most powerful model, but about choosing the one that provides the best response efficiently.

The Real Divide: Large Models vs. Small Models

The discussion is not limited to the opposition between proprietary models and open models. The real distinction lies between large and small models. Switching from GPT-5.5 to DeepSeek's V4 Flash may generate savings, but opting for GPT-5.4-mini can be just as effective.

An Ongoing Price War

There is active competition between the internal inference of large labs and independently served open-weight models. In the debate of small versus large models, it doesn't matter which type of small model prevails.

A Necessary Change in Mindset

While it may seem obvious not to use more resources than necessary, this idea contradicts the "scalability first" approach that has dominated thus far. Labs have long sought to develop the most resource-intensive models, pushing the limits of AI. With prices often subsidized by investors, customers had no reason not to choose the most advanced option.

Cost Pressures and Uncertain Future

With rising token prices and decreasing subsidies, users are feeling cost pressures for the first time. It remains to be seen whether this will actually push companies to adopt smaller models. They may also choose to reduce the number of calls, use less context, or abandon certain deployments.

If most deployments can function just as well with a smaller model, this could dampen the growing demand for inference and raise new questions about the justification of costs associated with training cutting-edge models.