OpenAI lost the plot on boring LLM use cases (2025)

I recently began updating my notebooks for Cheat at Search with LLMs .

It’s a training course using LLMs in what’s become their “boring” applications - NLP. I care about taking search queries and classifying them into high-level categories (furniture or electronics). Or extracting entities like colors or brands. Or correcting spelling mistakes. These used to be hard NLP problems that took months, now LLMs help practitioners like me quickly get to an 80% solution.

Back in my day (ie 2020), one the “cool” applications of LLMs - at the time we just said transformer models - was this zero/few shot classification. Now I realize the cool/sexy topic has become agents.

With GPT-5 OpenAI seems to have forgotten NLP use-cases in favor of the latter. Agents benefit from reasoning models. Boring NLP tasks want speed, low cost, and don’t benefit from time consuming reasoning.

With the NLP use-cases I care about, I was at one time enthusiastic about smaller open source models. They solved these problems well enough. They let me host models wherever I wanted.

Then the major providers, like OpenAI, released mini and nano models, and I realized it wasn’t “open source” I cared about, but just reasonably small and cost-effective. I care about structured outputs - specifically the ability to constrain to a small set of labels.

It turned out, I’m happy to pay a nominal fee for someone to host a tiny LLM for me to batch process thousands of search queries or documents at reasonable cost.

With this setup, I felt I had an ability to carefully dial my token usage, latency, and accuracy for classification problems I would never have dreamed of.

More importantly, I wanted to support OpenAI, who started this revolution. I want to be a fanboy.

Imagine my surprise when I started migrating my classification code to GPT-5.

They force reasoning into my carefully tuned classification pipeline. The best I can do is set reasoning=minimal . I can’t turn it off. And OpenAI doesn’t seem to have a roadmap for a model without “reasoning” (gpt-5-chat doesn’t have reasoning, but its targeting conversation).

Now I am an avid user of agents like Claude Code. And I realize reasoning carefully about evidence is important for these use cases.