Plan with the Expensive Model. Build with the Cheap One.
T007 · by Alexey Makarov — AI Strategist, Practitioner & Educator · Founder, AI Center of Excellence
When a new frontier model lands, don't ask whether it's worth the tokens — use it once, for the part only it can do (finding the solution you hadn't considered), and make it hand you a plan a cheaper model can execute.
Transcript
Read every word ↓Hide ↑
OpenAI releases new model, Anthropic releases new model, Grok releases new model. The question is, should you use the latest version of the model as of today immediately? Or maybe you should stick to the previous versions because they are less expensive and you'll burn your tokens not as quickly as with the new model. So my rule of thumb is the following. Once the new model is released, and I'm following this rule of thumb for the last, I think, six months, once new model is there, I try this new model on my project.
And I try it with more or less the following prompt. Please scan the repository of the project, check the implementation we are currently having, and try to come up with a solution which is innovative, which we have not considered, and which is significantly better than the one which we're having in the project. So the new model scans the repository, and normally it comes up to the solution, which I was not even considering. Then thinking about it, I'm saying, OK, that solution is actually quite nice, and it even implements the solution for me, normally in one go.
So for sure, later I can actually switch to the models which are not as expensive as the latest model from OpenAI, Anthropic, or any other company. But what I always do when I ask the model to come up with this innovative solution, I also ask it in the end to create a plan which can be used by a less capable model and can be implemented from the beginning till the end. So rule of thumb is the following. Once the new model is there, let's say GPT-6 is there, use it to come up with different innovative solutions about which you are not thinking about.
Ask it to analyze this project and then create a plan for a less capable model, and then you can switch to the previous versions of Claude models, previous versions of GPT models or the models in Cursor, and the plan will be implemented with this less capable model. So always use it. That's what I do, and I also recommend you to do it the same.
More Thoughts on AI
Newsletter
Found this useful?
Join the newsletter. You get one e-mail when something new and useful is out: a video, a guide, a tool. No spam, no schedule, and you can leave with one click.
AI Strategist, Practitioner & Educator. Founder & Global Lead of the AI Center of Excellence at SEFE. Creator of the Unreasonable AI YouTube channel. Based in Berlin.
About Alexey →