Can Cheaper AI Models Shake Up The Industry?

The AI revolution has long relied on a simple premise: Bigger is better, and the most advanced models reign supreme. Now, as mounting costs force companies to reassess their choices, it's unclear if smaller, more affordable options will disrupt this status quo.

From Big to Small

A recent prediction by Coinbase co-founder Brian Armstrong suggests that within 12-18 months, up to 80% of AI workloads could be handled by models costing 99% less. This shift would mark a significant change in the economics and competition landscape of artificial intelligence.

For years, companies have prioritized quality over cost, defaulting to the most advanced available model. But if these same tasks can now be accomplished with cheaper alternatives without sacrificing performance, it could lead to a seismic transformation in how AI is deployed and managed.

The Shift in Economics

One major impact of this shift would be financial. Big labs like OpenAI and Anthropic are currently facing the consequences of rising costs as they head toward their IPOs. If cheaper models can deliver comparable performance, these companies could face a significant financial challenge.

However, there's hope that companies might adapt by optimizing their usage, reducing calls, or even scaling back less promising deployments. The real question is whether users will embrace smaller models for most tasks and only use more advanced options when necessary.

Testing the Waters

Initial tests are showing promise. In a recent experiment by legal AI tool Harvey, they managed to reduce inference costs by 300% without compromising quality. The key was combining Claude Opus with Fireworks' GLM 5.1 for heavy tasks and using Opus for less intensive work.

Co-founder Gabe Pereyra of Harvey explained: 'Quality comes first, but the definition of quality is evolving from simply using the most powerful model all the time to finding the best model that gets the job done efficiently.'

The New Divide

This trend isn't just about proprietary versus open models. The real battleground is between large and small models. Users can save money by switching from GPT-5.5 to DeepSeek's V4 Flash, but a smaller version of an advanced model like GPT-5.4-mini works just as well.

The competition for inference services is heating up, with both in-house solutions from big labs and independently hosted open models competing aggressively on price.

A New Paradigm?

While the idea that you shouldn't use more compute than necessary might seem obvious, it challenges the scaling-first approach that has dominated AI development. Labs have pushed the limits of model size, driven by investor support and a lack of cost pressure from users.

Now, as token prices rise and subsidies slow down, companies are finally facing real financial constraints. This could lead to a significant reduction in demand for advanced inference services, raising questions about the value proposition of training cutting-edge models.