The great AI downsizing: Why cheaper models are suddenly the smartest bet

The artificial intelligence industry has long operated on a simple, powerful premise: bigger models are better, and the best model wins. This assumption has fueled a race for scale, with companies like OpenAI and Anthropic pouring billions into training ever-larger frontier models. But a quiet, potentially seismic shift is underway. Mounting costs are forcing enterprises to reconsider their reliance on the most expensive AI, and a new era of cost-conscious model shopping is beginning. The question is no longer just about raw power, but about efficiency, and the answer could reshape the entire AI economy.

The scaling assumption under pressure

For years, the AI industry’s trajectory was defined by the ‘bitter lesson’: that leveraging massive computation was the surest path to better performance. Labs competed on quality, which meant defaulting to the most advanced model available. Investors subsidized the high costs of inference, giving users little incentive to economize. Now, that dynamic is changing. Token prices are rising, subsidies are slowing, and enterprises are feeling real cost pressure for the first time. The natural response is to start shopping for cheaper alternatives.

Coinbase’s Armstrong predicts a dramatic shift

Coinbase co-founder Brian Armstrong has offered a stark prediction: within 12 to 18 months, 80% of AI workloads will run on models that are 99% cheaper than today’s frontier systems. Only the remaining 20% of tasks, those requiring maximum intelligence, will continue to use the latest generation models. If this forecast holds, it represents a fundamental change in the economics of AI. Much of the savings would come directly out of the revenue streams of major labs like OpenAI and Anthropic, potentially dealing a significant financial blow as they approach their IPOs.

Real-world tests show promise

Initial evidence suggests Armstrong’s prediction is not far-fetched. A recent test by the legal AI tool Harvey, conducted in partnership with the inference platform Fireworks AI, demonstrated that costs could be reduced by three times without any loss in quality. The system intelligently routed simpler tasks to a smaller, cheaper model (Fireworks’ GLM 5.1) and reserved the more powerful Claude Opus for the most demanding legal work. Harvey co-founder Gabe Pereyra noted that the definition of quality is evolving from simply using the most powerful model for everything, to using the best model that gets the right answer most efficiently.

The real divide: large vs. small, not open vs. closed

The emerging cost war is often framed as a battle between proprietary models from US labs and open-weight models from Chinese firms like DeepSeek. However, this framing misses the larger point. The critical divide is between large models and small models. A company can save money by switching from a frontier model to a cheaper open-weight alternative, but it can achieve similar savings by switching to a smaller, cheaper version from the same lab. The price war is between large-scale inference and small-scale inference, and for the broader industry shift, it doesn’t matter which type of small model wins.

What this means for the industry’s future

If most enterprise deployments can be run just as effectively on smaller, cheaper models, it would put a serious damper on the growing demand for inference. This, in turn, would raise difficult questions about how to justify the enormous cost of training a frontier model. The industry is at a crossroads. It could either embrace efficiency and risk slowing the growth of its most expensive products, or it could find new ways to demonstrate that the extra cost of a frontier model is justified. The answer will determine the winners and losers in the next phase of the AI revolution.

Conclusion

The AI industry’s foundational assumption is being tested. As enterprises face real cost pressures, the shift to smaller, cheaper models is no longer a theoretical possibility but a practical necessity. The impact could be profound, potentially slowing the revenue growth of major labs and forcing a re-evaluation of the entire scaling paradigm. The coming months will reveal whether the industry can learn to love cheaper AI models, or whether the demand for frontier intelligence remains insatiable.

FAQs

Q1: Why are cheaper AI models becoming more attractive now?
Rising token prices and a slowdown in investor subsidies are creating real cost pressure for enterprises that use AI. This is forcing them to look for more efficient options instead of defaulting to the most powerful model.

Q2: Will using cheaper models mean lower quality results?
Not necessarily. Early tests, such as the one conducted by Harvey, show that by intelligently routing tasks, companies can achieve the same quality while significantly reducing costs. The key is using the right model for the right job.

Q3: How would this shift affect companies like OpenAI and Anthropic?
A widespread move to cheaper models could reduce demand for their most expensive inference services, potentially impacting their revenue as they prepare for public offerings. It would challenge their business models, which are built on the assumption that customers will pay a premium for the best possible intelligence.