The video explains how model routing—using expensive AI models for planning and cheaper ones for execution—can cut AI costs by over 60%, exemplified by switching from a single frontier model to a combination of models for coding tasks. It also highlights practical implementation methods and encourages adopting this strategy broadly to optimize AI spending and productivity.
The video discusses a highly effective strategy called model routing to significantly reduce AI costs, potentially saving over 60% on AI bills. Model routing involves using different AI models tailored to specific tasks rather than relying on a single, expensive frontier model for everything. The key distinction is between planning and execution phases: planning requires a high-quality, expensive model to design and architect features, while execution—such as writing code—can be handled by cheaper, less powerful models that are still competent enough to perform the task well.
The presenter explains their workflow, starting with research and planning using a top-tier model like Fable, which produces detailed specifications for the feature to be built. Once the spec is ready, it is handed off to a more affordable coding model, such as GPT 5.5 or Composer 2.5, to write the actual code. After coding, the expensive model can optionally review the pull request to ensure quality, providing feedback that the cheaper model can then address before deployment. This back-and-forth approach leverages the strengths of different models while minimizing costs.
A detailed cost comparison highlights the savings: using a frontier model like Fable for both planning and coding can cost around $9.50 for a feature, whereas offloading coding to a cheaper model increases coding cost slightly but overall reduces total cost to about $3.20, resulting in approximately 68% savings. This is because planning uses fewer output tokens, which are more expensive, while coding involves many output tokens, so using a cheaper model for coding drastically cuts expenses.
The video also covers practical implementation methods, including manual switching between models using multiple subscriptions or automated workflows where models can call each other, such as Claude calling Codex. Additionally, third-party platforms like Cursor and Factory offer model routing features that automatically assign tasks to the most cost-effective model. These platforms support a variety of models, including open-source ones, providing flexibility and further cost efficiency, especially for enterprise users with large AI budgets.
Finally, the presenter emphasizes that model routing is not just for coding but applies broadly to all AI tasks. Users should familiarize themselves with different models’ strengths, costs, and “thinking” effort levels to optimize usage. Large companies like Coinbase have successfully implemented model routing combined with caching and context management to keep AI costs flat despite increasing usage. The video encourages viewers to adopt model routing strategies to maximize their AI investment and improve productivity.