How I view LLMs as Sept 2026

A developer who has spent over $30,000 on tokens across Codex, Claude, and GLM since the start of the year argues that large language models now fall into two categories: instruction-following models like Luna and Sonnet, which are cheap workhorses that lack common-sense judgment, and intent-understanding frontier models like Astra and Fable, which better grasp the intent behind a question and can orchestrate the cheaper models. The developer contends that today's models remain far from making human-like trade-offs, and that closing that gap is the path to AGI.

How I view LLMs as Sept 2026 I’ve spent over $30k in tokens since the start of the year across Codex, Claude, and GLM. Here’s how I view LLMs as of September 2026. Two categories: Instructions-following models: - Like Luna and Sonnet. - I used to call these “dumb”. But that’s the wrong way to think about them. - They just cannot make good “common sense” decisions. - They’re workhorses for pennies. Luna especially, is hard to spend a lot on. - If you make the core decisions up front, coding is auto-complete. Intent-understanding models: - Frontiers like Astra and Fable. - These are higher “common sense” models. These models try to understand the intent behind the question. - Solving problems is about making the right tradeoff - these models tend to make the right tradeoff more often than the other models. - These models know how to orchestrate instruction-following models prescriptively - in a way humans get too lazy to. In my experience, today’s models are far from making the right human-like trade-offs. However, this is where I believe the path to AGI is. That is, if a model can make the tradeoffs a human would.