Beyond the 'best' AI model: it's time to build your stack

The idea of a single 'best' AI model is outdated. A recent discussion around a viral AI model tier list highlights a more sophisticated approach: building a model stack. This involves strategically combining high-performance proprietary models for complex tasks with faster, cheaper open-source models for routine ones. For businesses, this method is key to optimising performance, managing costs, and unlocking practical AI adoption.

The AI community's focus is shifting away from crowning a single 'best' foundation model. As highlighted in the AI Daily Brief, the conversation now revolves around a more nuanced, multi-faceted evaluation of AI models, acknowledging that different tasks require different tools.

This new approach moves beyond judging models on raw intelligence alone. Instead, it balances three critical factors:

• Performance: The model's reasoning, accuracy, and creative capabilities.
• Speed (Latency): How quickly the model generates a response.
• Cost: The price per token or inference, which can vary dramatically.

This has given rise to the concept of a model stack (or 'model cascade'), where businesses use a portfolio of models and route tasks to the most appropriate one based on these criteria. A task might start with a fast, cheap model, and only escalate to a more powerful, expensive one if the initial attempt fails or the task requires deep reasoning.

A tiered view of AI models

We can categorise models into tiers to help decide which one is right for a given job. There is no single universal ranking, but the tiers generally follow this structure:

The most advanced organisations are no longer asking "Which model should we use?" but rather "Which combination of models provides the best overall performance and efficiency for our workflows?"

For any business integrating AI, moving from a single-model mindset to a model stack strategy has significant implications for productivity, cost, and risk management.

Cost optimisation is now achievable

Using a top-tier model like GPT-4o for every single task is like using a sledgehammer to crack a nut—effective, but incredibly inefficient and expensive. Many business tasks, such as classifying customer support tickets or extracting invoice data, do not require frontier-level intelligence.

By routing the majority of these high-volume, low-complexity tasks to faster, cheaper models, an organisation can dramatically reduce its AI operational costs. The expensive, powerful models are reserved for the small percentage of tasks that truly need them, maximising return on investment.

Better performance and user experience

Speed is a critical feature. For customer-facing applications like chatbots or internal tools used by staff, low latency is essential for a good user experience. A two-second delay can be the difference between a tool that feels seamless and one that feels broken.

A model stack allows you to use lightning-fast models for real-time interactions, ensuring the user gets an immediate response, while still having the ability to call on a more powerful model for deeper, asynchronous analysis when needed.

The foundation for effective agentic workflows

Agentic AI systems, which complete multi-step tasks autonomously, are rarely powered by a single model. An effective agent is a master of delegation. It might use a small, fast model to analyse an initial user request, a different model to search a database, and a third, powerful model to synthesise the findings into a final report.

Building a model stack is a prerequisite for building sophisticated AI agents. It provides the agent with a toolkit to intelligently manage its own computational resources, choosing the right tool for each sub-task in its workflow.

Reduced vendor lock-in

By building workflows that can leverage multiple model providers—including open-source alternatives that can be hosted on your own infrastructure—your organisation becomes more resilient. If a single provider has an outage, raises prices, or changes its service, your operations are not entirely dependent on them. This strategic flexibility is a crucial aspect of long-term risk management in the AI era.

Adopting a model stack approach requires a deliberate, strategic shift. It doesn't have to be an all-at-once transformation. Here are the practical steps to get started:

1. Audit your current and planned AI use cases. Go through your workflows and categorise the types of AI tasks you perform. Distinguish between simple tasks (e.g., data extraction, classification, formatting) and complex ones (e.g., strategic analysis, long-form content generation, multi-document synthesis).

2. Map tasks to model tiers. For each task category, identify the lowest-cost model tier that can reliably accomplish it. Create a simple matrix that assigns 'Frontier', 'Workhorse', or 'Specialist' models to your list of tasks. Your goal is to default to the cheapest, fastest option that works.

3. Implement a basic 'cascade' logic. Start simple. In your application logic, first send a request to a cheap, fast model. Check the output for quality or a confidence score. If it's not good enough or the model indicates it cannot handle the request, then escalate the same request to a more powerful, expensive model. This is your first model cascade.

4. Explore router-based systems. As your needs mature, investigate using a 'router' or 'gateway'. This is a dedicated component in your AI stack that directs incoming requests to the most appropriate model. The router itself can be a simple rules engine or even a small, specialised AI model trained to select the best model for a given prompt.

5. Measure, monitor, and refine. Continuously track the cost, latency, and quality (e.g., user satisfaction, error rates) for each model in your stack. Use this data to refine your routing rules. You might find that a model you thought was a 'Workhorse' can handle tasks you previously assigned to a 'Frontier' model, presenting a new opportunity for optimisation.