7 Oct 2026
Four AI models, one real task — the model-selection matrix
Over the last few weeks I compared four AI models on a real task for a client: suggesting restock orders for a wholesale distributor, around 560 products from two suppliers. I expected the comparison to be about which model is smartest. It ended up being about other things.
Three of the four models reached the same quality. I checked them on 13 products with tricky patterns, like bulk packs, sudden jumps in demand and slow movers. GLM-5.3-Flash and Claude Sonnet 5 caught all 13, and DeepSeek caught 11. Gemini Flash-Lite caught none. 98% of its answers were identical to the simple formula I had given it as a reference, so it added nothing.
With quality equal, the decision came down to cost, reliability and data terms. Sonnet cost about 17 times more per run than GLM for the same result. DeepSeek has one of the lowest prices per token, but it used about six times more output than Gemini, needed 25 retries, and one batch never finished. We chose GLM, after the client confirmed she was comfortable with her stock and sales data being processed by that provider.
I also wanted a carbon column. None of the four providers publishes energy or emissions per request, so I estimated it myself, and the ranges overlap too much to rank the models. I kept the column anyway, mostly to show the gap.
What surprised me most came at the end. When we compared the suggestions with the order the client actually placed, the differences came from the data, not the model. Products that had been out of stock showed no recent sales, and no model could tell that apart from no demand. For a task like this, I now think the model you pick matters less than the data you give it.
Originally posted on LinkedIn ↗