The most expensive mistake in agentic systems right now is not using too little AI. It is routing every task to the frontier model by default.
At Avakata, 82% of the tasks our engine ran in June went to small models — the cheap, fast tier that costs a fraction of a cent per call. Quality did not drop. The review gate caught the same number of defects it always catches. The bill fell 71%.
Here is how we decide which model gets which job, and why the cheap model is the right call more often than most operators believe.
What counts as a small model in 2026
A small model is any model that costs under roughly a tenth of a cent per typical task and returns in under a second — the compact tier every major lab now ships alongside its flagship. It will not plan a campaign or untangle an ambiguous brief. It will classify, extract, reformat, and summarize with near-frontier accuracy at 1-3% of the price. That price gap, not the capability gap, is the number that should drive your architecture.
The gap has widened all year. Flagships got smarter and slightly cheaper. Small models got dramatically better at the narrow tasks that fill an agent's day.
And most of an agent's day is narrow.
Names change quarterly, so we do not pick favorites in print. The tier is what matters: the fastest, cheapest model your provider ships that still follows instructions reliably.
The jobs where the cheap model wins outright
Small models win outright on tasks with a verifiable right answer and a narrow output format: classifying an inbound email into one of six buckets, extracting a date and an amount from an invoice, cutting a title to under 60 characters, tagging a support ticket, deciding whether a page changed enough to recrawl. In our June logs those five task families were 64% of all model calls, and the small model matched the flagship on 99 of 100 sampled outputs.
The pattern behind the list: bounded input, bounded output, cheap verification. If a regex or a schema check can confirm the output, you do not need a genius to produce it.
Formatting work is the clearest case. Rewriting a meta description to 155 characters is not thinking.
It is typing with constraints.
Summarization sits on the boundary. A neutral recap of a call transcript is small-model work. A summary that must decide what mattered belongs a tier up. We split those two jobs on purpose.
Where the frontier model still earns its price
Frontier models earn their price on tasks where the failure is expensive and hard to detect: strategy memos, first drafts of position pieces, ambiguous client emails, anything that requires reading between the lines or holding twelve constraints at once. In our stack that is roughly 18% of calls but 55% of spend, and we consider it money well spent. A wrong classification costs a retry. A wrong strategic read costs a client.
We also keep the flagship on anything customer-visible that ships without a human glance. One embarrassing paragraph outweighs a month of routing savings.
The rule is asymmetry. Pay up where mistakes are silent and costly. Pay down where they are loud and cheap.
Price the downside in dollars and the choice usually makes itself. Our worst small-model miss in June cost 40 cents of retries. Our worst plausible flagship-task miss would have cost a five-figure engagement.
Route by failure cost, not task difficulty
The routing question is not whether the small model can do the task. It is what happens when the small model gets the task wrong, and whether you will notice. Cheap-to-verify tasks with visible failures can run on the smallest model you have, because your checks catch the misses. Expensive, silent failures justify the flagship even when the task looks easy. Difficulty is a bad proxy. Detection is the real variable.
This flips some intuitions. Sentiment tagging looks trivial. But a mistagged angry customer who waits three days for a reply is a silent, expensive miss.
So that one runs on the mid tier, with a confidence threshold that escalates anything ambiguous.
Write the failure mode down before you route the task. If you cannot name the failure, you cannot price it, and the default should be one tier up until you can.
The escalation ladder we run
Our engine runs a three-rung ladder: the small model attempts the task first, a validator checks the output against a schema or rubric, and anything that fails twice escalates to the mid tier, then to the flagship. In June, 88% of ladder tasks resolved on the first rung, 9% on the second, and 3% reached the top. The blended cost per task came out 84% below flagship-only pricing on those pipelines.
The ladder only works because escalation is automatic. The moment a human has to decide which model to use, you have spent more attention than the routing saves.
Route with code, not with judgment calls.
One implementation note: log every escalation with the failing output attached. That log is where you discover which prompts are weak and which tasks were misfiled in the first place.
What the June bill actually looked like
Concrete numbers from our own engine: 412,000 model calls in June. Flagship-only pricing would have cost about $2,300. The routed bill was $660 — 71% lower — and the human review queue caught 41 defects, inside the same 38-to-45 band it has held since March. Cheaper did not mean sloppier, because the defect rate lives in the review gate, not in the model choice.
For a solo operator running a lighter stack, the ratios hold at smaller scale. A $180 month becomes a $50 month.
Latency budgets follow the same shape. Our daily content run dropped from 34 minutes to 9 when the extraction steps moved down a tier.
The savings are not even the whole story. Small models return in 300 to 600 milliseconds, so the pipelines feel different. Retries are painless. Batch jobs finish before the coffee does.
How to start tiering this week
Start with a one-hour audit: pull last week's model calls, sort by volume, and mark every task family that has a bounded output and a cheap check. Move the top three families to a small model behind a validator, and log every escalation. Most teams find that 50-70% of calls move safely on the first pass, and the logs tell you within a week whether quality held.
Do not tier everything at once. One family at a time, with a rollback switch, beats a grand migration.
Keep the validator honest, too. A check that passes everything is not a check.
Expect two surprises. A task you were sure needed the flagship will run clean on the small model for a month straight. And one trivial-looking task will escalate constantly until you rewrite its brief.
The frontier model is a specialist. Stop paying specialist rates for typing.