These include a gap between current model capabilities and the software products built to use them, as well as the immense capital expenditure required to train increasingly larger systems.
This shift in strategy matters because the competitive landscape is moving from pure model performance to product integration and infrastructure efficiency.
According to analysis from Stratechery, current models like Claude and GPT are already "good enough" for many enterprise tasks, leading customers to prioritize data privacy and cost over incremental intelligence gains.
Furthermore, competitors like Meta have demonstrated that slightly less powerful models can still power highly effective personal agents, suggesting that the "moat"—or defensive business advantage—of having the absolute fastest model is shrinking.
Slowing development could allow labs to address a significant "pricing overhang" by reallocating expensive computing power from training new models to serving existing ones more cheaply.
This would help companies like Anthropic and OpenAI reach true profitability by reducing the massive cash burn associated with research and development.
However, critics argue that pacing the frontier could actually increase safety risks, such as cybersecurity threats, by delaying the development of automated defensive tools needed to counter AI-driven attacks.