Every week now, someone's publishing a benchmark showing a local model that matches or beats a frontier model on some narrow task. The YouTube titles are relentless: "This model runs FREE on your laptop." "90% faster. Zero API costs." "Replace Claude for half your workflows."
And every week, someone's team hears this and starts routing real client work through whatever the free tier offers.
That's where it falls apart. Not at the benchmark stage. Not in the demo. At the moment someone needs to prove what actually ran, what was checked, and what would happen if the output was wrong.
The short version
Cheap AI models are useful worker lanes. They are not quality systems. A serious AI operation needs to know which work can run cheaply, which work needs stronger review, and how to prove the difference after the fact.
The problem is not that cheap AI doesn't work. The problem is that working and being accountable are different things.
Here's what we see constantly: a business drops a free or cheap model into a workflow that used to cost them real money. The output looks fine. The client asks what model was used. The team has no log. The client asks what the fallback was when it hallucinated a citation. The team points to a benchmark that doesn't apply to their specific task. The client stops asking questions but starts asking for someone else.
Cheap models are a worker lane. Not a quality system.
The honest version of tiered routing
The honest version of tiered AI routing looks like this: simple drafts, classification tasks, background data processing, and first-pass summarisation can run on local or cheap API models and save real money.
But each task needs three things before it goes anywhere important: a source, a check, and a known fallback.
If the cheap model produces output that matters, such as a proposal, a client-facing summary, or a claim about performance, it needs a human or a stronger model in the loop before it leaves the building.
The "free forever" people are not lying. Local inference and smart routing genuinely work for bounded tasks. But they are describing the cost layer, not the quality layer.
And quality is where most AI spend actually goes: the evals, the review, the approval, and the log that proves the work was done right.
Cheap is not the strategy
The companies running AI well right now are not the ones using the cheapest models. They are the ones who know exactly which tasks can run on cheap models, which cannot, and how to tell the difference before the work goes to a client or a public-facing surface.
That's not an intelligence problem. That's an operations problem. And it doesn't get solved by switching to a cheaper model.
It gets solved by building the tiered stack: cheap workers where you can verify output, expensive judgement where you cannot afford to be wrong.
If your team is routing anything important through a free tier without a log, a check, and a fallback, the model being cheap is the least of your problems.
What should an AI routing stack prove?
A serious AI routing stack should prove what model ran, what source material it used, what checks were applied, who or what approved the output, and what fallback exists when the task is too risky for the cheap lane.
When should businesses use cheaper AI models?
Use them for bounded, checkable work: first drafts, internal summaries, classification, data preparation, low-risk rewrites, and repeatable background processing. Do not use them as the final authority on claims, strategy, client recommendations, financial analysis, or anything where being wrong creates reputational risk.