---
type: idea
title: Do we need frontier intelligence for 80 percent of knowledge work?
created: 2026-07-07
status: seed
origin_sources:
  - LLM Evals ambiguity judgment experiment
related_auto: []
used_in:
  - "[[Do We Need Frontier Intelligence for 80 Percent of Knowledge Work]]"
---

## Claim

Maybe the question is not whether open source or lower-cost models are as good as frontier models in some abstract sense. Maybe the better question is whether we actually need frontier intelligence for 80% of knowledge work.

Frontier models are the obvious default when the work is high-stakes, open-ended, long-horizon, or hard to specify. GPT and Claude are broadly capable enough that they often absorb messy context and still give you something useful. That matters. But a lot of knowledge work is not pure frontier cognition. It is repetitive, bounded, domain-shaped work: reconciling data, drafting summaries, checking assumptions, cleaning names, preparing board notes, updating forecasts, or turning messy inputs into a structured artifact.

For that kind of work, the real question is not "which model is smartest?" It is: can a cheaper model make the right business assumptions, follow the workflow, avoid overclaiming, and produce something a human operator would trust?

## Why this matters

The cost of intelligence is becoming a real business problem. If every workflow routes to the most expensive frontier model by default, then AI adoption becomes economically lazy. It may work, but it does not teach the organization anything about where intelligence is actually needed.

The future probably looks more like model routing than model worship:

- frontier models for the hardest 20%;
- cheaper capable models for repeatable 80%;
- deterministic software wherever the answer should not be probabilistic at all.

But you only get that architecture if you can measure the work. Otherwise everyone argues from vibes.

## Evidence / examples

Over the weekend I ran a small finance eval to test this directly. I compared GLM 5.2 through a local relay with default Codex on five ambiguity-heavy finance tasks:

- stale source labels;
- split customer payments;
- dirty customer names;
- board-pack driver discipline;
- preliminary close / partial cutoff language.

These were not benchmark trivia questions. They were tiny versions of real finance work where the prompt is not perfect and the model has to infer what a finance person would do.

The result was more nuanced than "frontier wins" or "open source wins." Once I fixed the relay issue and corrected a too-literal grader, both models passed all five cases. GLM even scored slightly higher in the clean run because it carried more numeric/source evidence into the output.

That is interesting. It suggests that at least some finance knowledge work may not require the most frontier model. It may require a model that is good enough, plus a good eval, plus a clear sense of where failure is acceptable and where it is not.

## Counterpoints

This does not mean frontier models are overkill everywhere. They are still the better default when:

- the task is poorly bounded;
- the cost of failure is high;
- the workflow requires long-horizon agency;
- the context is broad and under-specified;
- the model needs to recover from bad instructions without help.

Also, my eval was tiny: five cases, one clean seed, hand-authored finance scenarios. It is evidence, not a universal law.

## Draft hook

If you have worked with large language models for a while, you probably have a default instinct: use the best frontier model available. That instinct is reasonable. GPT and Claude are polished, broadly capable, and forgiving. They are what you reach for when you want fewer surprises.

But defaults are expensive. And as open-weight and lower-cost models improve, the more interesting question is no longer whether they are "as intelligent" in some universal benchmark sense. The question is whether they are good enough for the work most people actually do.

So I spent a weekend testing that question in a narrow domain I understand: finance workflows.

## Possible headline

Do We Need Frontier Intelligence for 80% of Knowledge Work?

## What this is really about

This is not an article about GLM beating Codex or Codex beating GLM. It is an article about replacing model vibes with workflow-level measurement.

The argument is that AI strategy should become less about picking one default model and more about building an intelligence allocation system.