The Impending Paradigm Shift: Can the AI Industry Embrace Cost-Efficient Models?

The AI boom has been predicated on a fundamental assumption: that larger models are inherently more powerful, and consequently, the most powerful models will prevail. However, the industry now stands on the cusp of discovering the ramifications should this assumption begin to erode. Mounting costs have already compelled users to reassess smaller, more economical models. This nascent cost-conscious model-shopping represents a departure from established norms, and while its ultimate impact on the industry remains uncertain, it is likely to be profound. A particularly salient prediction, articulated by Coinbase co-founder Brian Armstrong, posits that the vast majority of tasks—specifically 80% of workloads—will migrate to models that are 99% cheaper within 12 to 18 months. The remaining 20% of workloads, he contends, will continue to rely on the latest generation models where maximising intelligence is paramount. The magnitude of such a shift, should Armstrong's forecast materialise, cannot be overstated. Historically, AI companies have competed on quality, which has invariably entailed defaulting to the most advanced model available. If cheaper models can indeed handle these tasks without compromising quality, the economic landscape of AI would be fundamentally altered. Critically, a substantial portion of the savings would be extracted from the revenues of major labs, delivering a financial blow to OpenAI and Anthropic precisely as they approach their IPOs. Preliminary tests lend credence to the viability of this transition. In a recent evaluation conducted by Harvey, a legal AI tool, the company managed to reduce inference costs by a factor of three without any degradation in quality. This test, executed in collaboration with the inference platform Fireworks AI, strategically combined Claude Opus with Fireworks' GLM 5.1, reserving Opus exclusively for the most computationally intensive tasks. The outcome was a markedly reduced load in terms of server time and overall expenditure. Gabe Pereyra, Harvey's co-founder, remarked that while quality remains paramount, its definition is evolving; it now encompasses using the most efficient model that yields the correct answer, rather than simply deploying the most powerful model indiscriminately. This trend is frequently contextualised as a conflict between major labs and Chinese or open-weight models, yet such framing obfuscates the more critical dichotomy. The genuine divide is not between proprietary and open models, but rather between large and small models. One can economise by transitioning from GPT-5.5 to DeepSeek's V4 Flash, but switching to GPT-5.4-mini proves equally effective. An aggressive price war is currently underway between in-house inference from the major labs and independently served open-weight models. For the overarching question of small versus large, the specific provenance of the small model is ultimately inconsequential. While this may appear self-evident, it fundamentally contradicts the scaling-first paradigm that has hitherto dominated the industry. Inspired by the bitter lesson, labs have heavily invested in training the most compute-intensive models possible. With prices heavily subsidised by investors, clients had no incentive to select anything other than the most advanced option. Now, as token prices rise and subsidies diminish, users are confronting cost pressure for the first time. Whether this pressure will indeed drive enterprise users toward smaller models remains to be seen. They might alternatively economise by reducing the number of calls, utilising less context, or abandoning the least promising deployments. Nevertheless, if it transpires that most deployments can operate just as effectively on a smaller model, this could significantly dampen the burgeoning demand for inference, thereby raising fundamental questions about the justification for training frontier models.