Create your own AI fusion model

Hi friend!
​
There's no shortage of new model releases these days. And all of the top AI models have different capabilities, strengths, and weaknesses. They come at different speeds and different price points.

Which one you should use depends not only on the task at hand, but also on the specific phase of the task.
​
Writing a good plan requires a different skill set than wiring up a service to an API. Gathering and synthesizing research is different from designing a visually compelling presentation.
​
Yet I currently see much more discussion about which model to use generally than about which model to use at each stage of a task’s lifecycle.
​
At the same time, the idea of model fusion is gaining traction. There are several interesting approaches emerging, including Sakana’s Fugu and OpenRouter’s Fusion.
​
Both combine the strengths of different models with some form of routing, promising outcomes comparable to top frontier models at a lower cost.
​
The cool thing is: with a bit of judgment and tinkering, you can effectively assemble your own fusion model.

For me, this often looks something like this:

• Let a top frontier model such as GPT 5.6 Sol, Opus 4.8 or Fable 5 write a tight plan (I usually use my own plan-writing skill for this)​
• Let a smaller, faster and cheaper, but still capable model write the code and run basic tests. My current favorite is Cursor's Composer 2.5, but i.e Sonnet will also work well.
• Bring in a larger, more expensive model again to run a thorough code-quality review (typically I use this skill)
• Let smaller models implement the fixes based on the review's findings.
• Loop the last two steps until all tests pass and the bigger model's code quality review comes back clean.

Depending on your needs, you can customize this further for example by using distinct models for i.e. code security review vs. design review.

There's no real right or wrong here - I think of big vs. small models as a directional choice. Which specific model you need when comes down to your personal taste and experience.

With model-specific subagents in tools such as Cursor, you can package the workflow above into predefined, reusable skills.

In other words, you can build your own fusion model.
​
Try it out and let me know what you think!​
​
​Thanks for reading,

Robert