Small Models, Big Jobs
Vishal Sachar
Co-Founder & CEO of CLRT
The reflex when the agent bill grows is to wait: prices fall, models improve, and next quarter's frontier will make today's economics look quaint. The reflex misses where the money actually goes. An agent run is not one act of intelligence. It is a long chain of model calls, and most links in the chain are mundane: classify this message, pull these fields, check that format, summarise this thread. A minority are judgment. Send every link to the flagship model and you are paying frontier prices for clerical work, on every run, indefinitely. The cost curve of agentic AI does not bend at the frontier. It bends at the router, the layer that decides which call deserves which model, and almost nobody engineers that layer on purpose.
Watch a production agent closely and the mystique dissolves into a call log. A run that looks like one act of intelligence is actually dozens of model calls, and most of them are clerical. Classify this message into one of nine categories. Pull the invoice number and the amount out of the attachment. Check that the output matches the schema. Compress this thread into four lines. These are jobs with known answers and checkable outputs, and they repeat identically on every run. Only a minority of calls carry the weight a frontier model exists for: planning the run, weighing an ambiguous case, deciding that something is an exception worth a human's attention. Yet the default architecture sends every call, clerical or not, to the most expensive model available, because a single model was the easiest thing to wire on day one.
The strongest statement of the alternative comes from an awkward source. In June 2025, NVIDIA Research published a position paper arguing that small language models are the future of agentic AI, and it put estimates on the claim. Examining three open-source agents, it judged that roughly 70 percent of Cradle's model calls, 60 percent of MetaGPT's, and 40 percent of Open Operator's could be reliably handled by small models, at what the paper puts at 10 to 30 times lower inference cost. This is a position paper from a vendor with an interest in every tier of the market, so read the numbers directionally. But the direction is hard to dismiss, and the floor is already remarkable: Stanford's 2025 AI Index records a capable small model priced at seven cents per million tokens. The question is no longer whether the routine majority of calls can move down a tier. It is whether your system can move them safely.
What happens next in most companies is routing by vibes. Someone sees the bill, swaps a cheaper model into the prompts that look easy, watches nothing break for a week, and declares the optimisation done. Nothing broke because nothing was being measured. Routing is an engineering discipline with named parts: an explicit map of which subtask runs on which tier, an evaluation suite per tier so that a downgrade is a measured decision rather than a hopeful one, a fallback for when the small model misses, an escalation path for the call that arrives harder than its category, and regression checks for the day a provider quietly updates a model underneath you. None of this is glamorous. All of it is the difference between a cost curve that bends and a quality curve that collapses in private.
The reason vibes fail is that the cheap tier fails differently from the expensive one, and more dangerously. A frontier model asked to do something hard fails in ways people are already watching for. A small model asked to do something easy fails plausibly, at scale, upstream of everything else. Classification is usually the first call in the chain, so a small model that is quietly wrong on four percent of categories does not produce four percent of bad labels. It produces four percent of runs in which every downstream step, including the frontier-model judgment steps, executed perfectly against the wrong premise. The failure surfaces far from its cause, wearing the costume of a good run. Catching it requires the per-tier evals and confidence signals designed before the downgrade, which is precisely the work that routing by feel skips, because the saving appears on the invoice and the quality debt does not.
Beneath the engineering sits the question that decides everything, which is where this stops being procurement. Deciding which calls are clerical and which are judgment is not a property of the model. It is a read on the workflow: what the output feeds, what a wrong answer costs, whether wrongness is even detectable before it reaches a customer. Two businesses can run the same nominal subtask, and in one it is safely small-model work while in the other it is the exact step where the money is lost. The frontier model is not the engine of an agentic system. It is the specialist called in at the steps that deserve it, and knowing which steps those are, in your business rather than in a benchmark, is the scarce work. The router merely enforces the answer.
The agent cost curve bends at the router, not at the frontier.
A deeper dive
The discipline has a shape, and it is worth seeing why it defeats casual attempts. Tiering only works when evaluation is the substrate rather than an afterthought: a small fixed test set per subtask, scored the same way every time, so that the sentence "the cheap model handles classification" is a measured claim with a number attached rather than a feeling that nothing has broken lately. That substrate is what makes everything else possible. Fallbacks need a definition of failure to trigger on. Escalation needs a confidence signal that is not the model's own self-assessment, because models grade their own work generously; it needs proxies such as schema violations, retrieval coverage, disagreement between two cheap calls, or a separate checker scoring the result below a threshold. And the whole arrangement needs maintenance, because the model roster underneath it changes every quarter: providers deprecate, prices move, and a new small model makes yesterday's routing map wrong in the profitable direction. A routing layer that cannot be re-evaluated cheaply is a routing layer that ossifies, and an ossified router quietly becomes the vibes it was built to replace.
There is also a compounding logic here that most teams miss. Chat is retail consumption of intelligence: one question, one answer, a human absorbing the cost of every mistake. Agents are industrial consumption: loops, retries, tool calls, calls that trigger other calls, running on schedules with nobody watching. In an industrial system, per-unit economics and per-unit error rates are the whole business, which is why serious industries meter their inputs and grade their outputs by stage rather than buying the premium grade of everything. The single-model agent is the equivalent of running a factory entirely on the most expensive input available because nobody built the plumbing to use anything else. The tell of an immature build is model choice as one global setting; the tell of a mature one is a routing map someone can defend line by line, with the evidence attached. Getting from the first to the second is not a configuration change. It is judgment about the workflow, encoded as engineering, and it is exactly the layer that decides whether agentic economics ever work in your favour.
Work with CLRT
Deciding which steps in your operation are judgment steps, and engineering the tiering, evals, and escalation that let everything else run cheaply and safely, is the work CLRT does. It is also the question our diagnostic starts with: where in your business does the frontier actually earn its price. If your agent bill is growing and nobody can show you the routing map, run the CLRT Ascent diagnostic at ascent.clrtstudio.com, or bring us the workflow and we will draw the map with you.
Vishal Sachar
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
