For years the AI industry operated on one core assumption: the biggest model wins. That belief is now breaking down. Companies are increasingly choosing models by task, cost, and control rather than benchmark position. The frontier still matters, but it is no longer the only thing being bought.
The reason is unromantic. At enterprise scale, model bills run into millions of dollars a month. The economics of using the largest, most capable models have become unsustainable for many businesses. This has forced a fundamental reconsideration of what constitutes value in AI deployment.
The rise of 'good enough' AI
The operating principle has shifted to the cheapest model that clears the quality bar. Buyers have worked out that most tasks do not need a frontier system. A summarisation job and a multi-step reasoning job no longer go to the same model. This segmentation is driving efficiency and cost savings.
Model routing has emerged as a key technology to automate that judgment, sending each request to whichever model suits it best. This allows enterprises to dynamically balance cost and performance without manual intervention. The approach is gaining traction across industries where AI usage is scaling rapidly.
Specialised models fill the gap
Specialised, industry-specific models are filling the rest of the gap. Instead of relying on a single monolithic system, businesses are deploying task-specific AI agents. Gartner expects 40% of enterprise applications to embed such agents by the end of 2026, up from under 5% a year earlier. This marks a dramatic shift from the one-size-fits-all philosophy that dominated earlier AI strategies.
These specialised models are trained on domain-specific data, allowing them to achieve high accuracy for narrow tasks without the overhead of a general-purpose frontier model. For example, a financial services firm might use a model fine-tuned for regulatory compliance while employing a separate model for customer service. The result is a portfolio of AI tools, each optimised for a particular function.
Why the bills forced this transition
The economics of large language models stopped adding up for many enterprises. Per-token prices have collapsed dramatically over the past year, yet enterprise AI bills have tripled anyway. This paradox is explained by the rise of agentic tools that consume vastly more tokens per task. Complex multi-step reasoning, long-context processing, and iterative refinement all drive up token usage exponentially.
Buyers noticed this trend. Palo Alto Networks chief executive Nikesh Arora has said that token prices need to fall by as much as 90% for adoption to scale meaningfully. Without such price reductions, many applications remain cost-prohibitive. Some firms gave up waiting and started rationing AI usage. A wave of 'tokenminimizing' has companies capping employee AI spending outright, limiting who can use the most expensive models and for what purposes.
This rationing reflects a broader concern about AI cost management. IT departments are increasingly implementing budgets and usage policies to prevent runaway spending. The era of unlimited AI experimentation is giving way to disciplined, ROI-based deployment.
Where the value moves next
If AI capability is commoditising, the margin migrates to whoever can run inference cheapest. Inference optimisation has quietly become one of AI infrastructure's most valuable layers. Startups and cloud providers are racing to develop more efficient hardware, better scheduling algorithms, and model compression techniques that reduce the cost per query.
Open and cheap models sharpen the competitive point. Chinese models, such as those from DeepSeek and others, are closing in on the capabilities of US frontier labs at a fraction of the price. This caps what anyone can charge for merely competent output. The commoditisation of basic AI capabilities is forcing companies to differentiate on cost, reliability, and integration rather than raw performance.
This is uncomfortable for the scaling thesis that has driven the industry. Hundreds of billions in capex were justified by the premise that bigger models would stay decisively better. Buyers are now voting otherwise, prioritising economy over power. The shift does not mean frontier models are finished. Research continues to push the boundaries of what AI can achieve. But it means the industry is discovering that most practical work is mundane and does not require the most expensive tool in the shop.
Enterprises are increasingly adopting a portfolio approach: a powerful model for complex tasks, a lightweight model for routine operations, and a suite of specialised agents for domain-specific work. This hybrid strategy optimises cost while preserving access to frontier capabilities when needed.
The implication for AI companies
For AI labs and model providers, the changing landscape requires new strategies. The race is no longer solely about model size; it is about efficiency, flexibility, and integration. Companies that can deliver fast, cheap inference will win enterprise customers. Those that focus exclusively on pushing the frontier may find themselves with a product that is technically impressive but commercially unviable for most use cases.
Additionally, the rise of model routing and agentic architectures creates opportunities for middleware platforms that orchestrate AI resources. These platforms will become increasingly important as enterprises seek to manage their AI portfolios effectively. The ability to automatically route requests to the most cost-effective model is a key value driver.
Another trend is the growing interest in on-premise and edge AI. To control costs and maintain data privacy, many companies are exploring running smaller models locally rather than calling expensive cloud APIs. This is leading to investment in hardware optimisation and model distillation techniques.
The regulatory environment is also likely to shape the market. As AI becomes more pervasive, governments may impose requirements on model transparency, safety, and energy consumption. Compliance costs could further incentivise the use of simpler, more efficient models that are easier to audit and certify.
Ultimately, the shift from model size to model economics represents a maturation of the AI industry. The hype around giant models is giving way to practical deployment considerations. This is healthy for long-term adoption, as it aligns AI capabilities with real business needs. The winners in the next phase will be those who understand that bigger is no longer always better.