On Tailored Software

When DeepSeek V4.1 Flash came out, I first tried it through an OpenCode Go subscription, since it was the simplest way in. Make an account, swipe my credit card metaphorically, log in through my harness, and I’m set. I liked the model and wanted to use it more, but I burned through that month’s limit in only a few days.

Until then all my personal AI usage had been subscription based, and we all hear how heavily subsidized subscriptions are. So I wanted to see how far the same $10 I spent on Go would have gotten me on the DeepSeek API directly. Two caveats first. About a third of my Go usage went to other models, and at the time OpenCode had a 4x limit promotion running for DeepSeek V4.1 Flash, which inflated what Go was giving me. Even so, the 632 million tokens I ran through DeepSeek would have cost only $6.61 at API pricing, and my cache hit rate on the API turned out similar to the 98.64% I saw on Go. With a third of the plan going elsewhere, that works out to about the same value as the $10 I paid, so even with the promotion inflating it, Go was not significantly better than the API. I will attach the generated report below.

The point of this is not to dive deep into the technical details of token billing on subscriptions and APIs. This cost breakdown, presented in a digestible way, compelled me to move my DeepSeek usage over to the API, which gives me faster speeds at about the same value, something I never would have expected. Had it not been for the breakdown, I would have been too lazy to run these numbers myself, and I would never have tried the API. By cost I mean a mix of my own human effort, real financial impact, and overall time spent. That cost is going down more and more as agentic capabilities increase, and not only for synthesizing data into silly little HTML files, but for all forms of software. Things that would previously be considered a hassle and not worth the effort are now worth trying. I only did this because the cost of doing it has gone down.

Maybe some software you use does a few things you wish it didn’t, or doesn’t do a few things you wish it did. Depending on its complexity, you may be able to create your own tailored version! I have already started doing this, and it was not difficult at all. Macasnap is one example. I was inspired by Omasnap, a screenshotting tool designed for Omarchy that was created by Shopify CEO Tobi Lutke. Most macOS options either cost money or weren’t quite what I wanted, so I made my own. I did the same with Splid: I built an expense splitter for just me and my girlfriend, with the things we actually need, like multiple currencies, cash withdrawals, recurring expenses, and OCR for receipts. The friction of using a less than ideal product used to be smaller than the effort of building a replacement, but that is no longer the case.

I think that custom tailored software with a userbase of n=1 has a strong future, for the reasons outlined above. Every feature is built for you, and nothing is there if you don’t want it to be. This will only become easier as time goes on. The awareness of these possibilities will diffuse through the general population, and the space of possibilities will keep expanding, leading to more and more people attempting things like this. I suspect that in the near future, many products, particularly quality-of-life apps, will take a hit to their user base as people build their own custom solutions.