The Position · Edition 4

Stop Swinging a Sledgehammer at a Thumbtack: Right-Sizing AI Models

3 July 2026

Every AI vendor conversation right now starts the same way: "We use Claude top class models for everything, so you know you're getting the best."

It sounds reassuring. It's also quietly becoming the most expensive mistake in enterprise AI procurement.

Here's the position: the era of the one-size-fits-all frontier model is ending — not because the models got worse, but because the economics stopped making sense, and the smartest CI and product teams are already moving to right-sized model architectures instead.

The subsidy is running out

For the past two years, frontier-model pricing has been artificially generous. Compute-rich labs have effectively subsidised inference to win developer mindshare and market share. That phase is closing. Frontier model costs are climbing as providers shift from land-grab pricing to margin-recovery pricing — and the vendors who built their entire stack on a single large model, for every task, are now passing that cost curve straight to customers.

The uncomfortable part: most of that spend was never buying better output. It was buying convenience.

Not every task deserves a frontier model

A large frontier model is genuinely worth its cost when a task requires broad reasoning, ambiguity resolution, or synthesis across loosely related information — the kind of work a senior analyst would do.

But a huge share of what agentic systems actually do, task by task, is narrow:

Classifying a competitor mention as "pricing," "feature," or "hiring" news
Extracting a job title and company from a leadership page
Summarising a single earnings call transcript into a fixed format
Tagging sentiment on a review
Structuring scraped text into a battlecard field

None of that needs a model that can also create a video, debug code, and reason about geopolitics. It needs a model that does one thing, fast, cheaply, and consistently — every single time it's called, at agent scale, thousands of times a day.

Running every one of those calls through a large frontier model isn't rigour. It's a big hammer looking for a thumbtack, billed at hammer prices.

Why "bigger" can even mean "worse" here

Counterintuitively, oversized models can underperform smaller, task-tuned ones on narrow work:

Latency compounds. Agent workflows chain dozens of calls. A large model's extra reasoning time multiplies across every step.
Over-generalisation creeps in. A model trained to handle everything sometimes "helpfully" reinterprets a rigid extraction task instead of executing it literally.
Consistency drops. Narrow, specialised models tend to produce more stable, repeatable structured output than generalist models asked to behave narrowly.

Right-sizing isn't a compromise on quality. Done properly, it's often a quality upgrade for the specific task — with a materially lower bill attached.

How we've built Pro-Position.AI around this

We don't route every agent task through a single expensive model, and we don't outsource that decision to a black box. Our architecture matches the model to the job:

Multiple open-weight models, each selected and tuned for a specific task class in the CI pipeline — extraction, classification, summarisation, synthesis
On-premise GPU infrastructure, removing per-token dependency on frontier-provider pricing for high-volume, narrow tasks
Large model reasoning reserved for the genuinely hard problems — cross-source synthesis, strategic framework generation, nuanced win/loss analysis — where that horsepower actually changes the output

The result isn't a trade-off between cost and quality. It's lower cost to serve and more consistent output on the tasks that make up most of an agent's actual workload — savings we pass through in price, not just margin we quietly keep.

The framework to steal

Before your next agent build, ask three questions per task:

Does this task require reasoning across ambiguous or loosely connected information? → Frontier model.
Is this task narrow, repeatable, and structurally defined? → Smaller, tuned model.
Is this task being run at high volume, many times per hour, per day? → Cost sensitivity should dominate the decision — and a right-sized model almost always wins.

If you're not asking these questions per task, you're not architecting an AI system. You're renting one, at whatever price the frontier lab sets next quarter.

Where this bites first: anyone whose entire CI or agent stack is a thin wrapper over a single frontier model API is exposed to every price increase that provider makes, with no lever to pull. That's not a technology risk. It's a vendor-concentration risk wearing an AI costume.

First published as an edition of The Position, Pro-Position.AI's newsletter on competitive intelligence, product strategy and go-to-market.

Keep exploring: