# Balancing Determinism vs. Non-Determinism and Build vs. Buy Through Doing the Work to Understand My Capabilities

**Published:** 2026-09-20  
**Author:** Kin Lane  
**Canonical:** https://apievangelist.com/2026/09/20/balancing-determinism-vs-non-determinism-and-build-vs-buy-by-understanding-my-capabilities/

It costs me an average of $2.00 to profile each provider listed in APIs.io. I have 50+ skills that I will run on each domain identified in the API Evangelist and APIs.io catalog. The catalog has ~29K providers in it, ~140K APIs, with ~650K operations, and many other supporting artifacts. That is roughly $60K I have spent this year to assemble the catalog. I have another ~15K providers in my harvest queue today, which will cost me about another $30K to pull off by the end of the year. It doesn't stop there. These are just the new providers, and as my profiling pipeline evolves, I need to revisit aging catalog listings.

My emotional reaction to this was that I need to train my own models and/or explore the usage of open-source LLMs to help me reduce my provider profiling costs. So I fired up the AWS infrastructure needed and got to work developing an artifact model that would do a lot of what I need with OpenAPI, AsyncAPI, JSON Schema, Arazzo, Overlays, JSON-LD, Plans, Rate Limits, APIs.json, and other essential artifacts that go into developing the API Evangelist and APIs.io catalog. I got my AWS SageMaker and other infrastructure up and running and began breaking down the training of the models, and how they would learn from the rich number of artifacts available across the API Evangelist GitHub repositories.

After spending the evening crafting a plan, and having Claude execute on the plan overnight, I woke up this morning with a working model I could use to create, manage, and evolve OpenAPIs—just one artifact in my catalog. Then I went for a three-hour walk through Central Park up to 81st for the Upper West Side Farmers Market, and back home again. Throughout the walk my brain rattled through all of the OpenAPI capabilities I am going to need, and then I began iterating through all of the capabilities I will need for the myriad of other artifacts. After the mental accounting, I determined that probably about 25% of my capabilities need to be non-deterministic, and the rest really are deterministic scripts that handle the inputs and outputs of each step. It is something I'll verify with more actual hands-on work, but once I got home, I rolled back all of the infrastructure for my model development and sat down to work mapping out my needs.

After writing a story yesterday on the [Frontier Storytelling](https://apievangelist.com/2026/09/19/frontier-storytelling/) around [TypeSafe's new System One Models and Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), I found myself thinking about how desperately we want our models to be structured and type safe. I can't help but think this emotion is healthy, but the scope of the toolbox we are looking for solutions in is too narrow. I think models have their place in our toolbox, but I think we are limiting ourselves when it comes to our obsession with AI in this moment. I need LLMs to craft summaries and descriptions for OpenAPIs, JSON Schema, and other artifacts. I also need LLMs to translate HTML API documentation into OpenAPIs, but there are a huge number of other things I do to these OpenAPIs that do not need to be done by AI. This is just one artifact. Most of the other things I need can be driven by code I generate and evolve using Claude, but there is no reason for me to be accomplishing each task with language models.

Breaking down each individual task within each individual skill in my provider profiling pipeline is a daunting task. They handle a lot of functionality. I pack a lot into that $2.00 profiling cost. It's worth it right now. But it is becoming a burden at scale, and it makes much more sense for me to optimize what I do, and how I do it. The impulse is to continue using AI, but develop my own models and/or use open-source models. I still have more exploration to do in using open-source models, but for what I do as the API Evangelist, the cost savings, performance gains, and quality improvements will occur within better mapping of what I need done, not falling into the continued trap of believing that AI will magically do it at scale. I need a detailed breakdown of my profiling pipeline, and what type of resource I need to accomplish each task. Some of those resources will be AI models, some will be 3rd-party APIs, and others will be my own APIs. For me the economics of this transcend this AI moment and go back to the traditional API economics of whether I build or buy, thoughtfully assessing each 3rd-party service based upon the value delivered, usage cost, and switching costs.

I don't doubt that at some point I still might craft my own models. I don't doubt that I may begin switching some capabilities over to open-source models. I don't doubt that I will keep using Claude and Gemini as core resources. But what will matter in the end is that I have a clear understanding of my capabilities and skills, and have them well crafted to help me optimize and route the technology and business of my APIs. Balancing the determinism vs. non-determinism and the build vs. buy of each individual operation that I execute. I think there is a lot of room for unpredictable outputs and costs without a clear assessment of your operations. Obviously this is how I approach my operations. Obviously this will vary across different operations. Obviously it will change as you scale operations. Obviously it will vary depending on what industries you operate within. Obviously it will vary depending on which country and region you operate in. There are no easy answers here, and it is clear that if you are looking for easy answers now, then you are just kicking the can down the road, and reducing the amount of control you have over your business in the short-term and long-term futures.