Cheaper AI Models Are Back: Why Model Routing Matters
As AI moves into real workflows, teams need model routing to balance cost, quality, speed, and human review across agent work.
Large AI models are expensive.
That is no longer just a complaint from finance teams. It is becoming a product design constraint for every company trying to put AI agents into real workflows.
On June 9, TechCrunch asked a direct question: Can tech companies learn to love cheaper AI models? The article points to a practical shift: companies still want the best models, but they no longer want to use the most expensive model for every step.
CNBC made a similar point on June 5, framing model routing as a response to AI overspending. Easy, high-volume work can move to cheaper models. Hard, risky work still goes to frontier models.
The story is not that strong models matter less.
The story is that strong models are becoming a scarce resource to route carefully.
What happened: cheaper models are becoming useful again
For the last two years, AI product positioning was simple: use the strongest model available.
That worked in demos. It works less well in production.
A real agent task is not one prompt. It can involve reading files, searching sources, planning steps, writing code, checking output, fixing mistakes, and producing a final artifact. If every step uses the most expensive model, cost rises quickly.
TechCrunch highlighted a test involving Harvey and Fireworks AI. According to the report, Harvey combined Claude Opus with Fireworks' GLM 5.1 and reduced inference cost by 3x without reducing quality. The key was not replacing the strong model everywhere. It was using cheaper capacity for the parts of the work that did not need Opus.
That is the important signal.
Cheaper models are not winning because teams stopped caring about quality. They are winning because teams are learning where quality is actually needed.
Why it matters: AI cost is now workflow cost
The cost question changes when AI leaves the chat box.
In a chat interface, one expensive model call may feel acceptable. In an agent workflow, the same task may trigger many calls: classify, summarize, retrieve, draft, execute, verify, revise, and report.
That turns model choice into workflow architecture.
Teams need to decide:
- Which steps are routine enough for cheaper models?
- Which steps need speed more than deep reasoning?
- Which steps affect customers, code, contracts, money, or production systems?
- Which steps require human approval before the agent continues?
- Which steps should be logged so managers can audit cost and judgment later?
This is why model routing is more than an engineering optimization. It is a product capability.
A product that always calls the strongest model may look good in a demo, but it can be too expensive to scale. A product that always calls the cheapest model may control cost, but it can fail exactly where judgment matters.
The hard part is the layer in between.
The new product question: who decides which model works?
A mature AI system does not ask one model to do everything.
It routes work.
Simple extraction can go to a cheaper model. High-volume classification can go to a fast model. Complex planning can go to a stronger model. A risky action can go to a stronger model and then stop for human confirmation.
This is similar to how a company works. Not every decision goes to the CEO. Not every task belongs to the intern. Good organizations assign the right level of judgment to the right job.
AI products now need the same discipline.
What teams should do next
Teams adopting AI agents should stop evaluating products only by asking which model they use.
That question still matters. It is just incomplete.
Ask these instead:
-
Does the system route by task difficulty?
Routine steps should not consume frontier-model budget by default. -
Does the system route by risk?
Customer-facing, financial, legal, production, or security-sensitive steps need stricter handling. -
Does the system explain model choice?
Managers should know why a step used a cheap model, a fast model, or a premium model. -
Does the system keep cost visible?
AI cost should be visible at the workflow level, not discovered after the bill arrives. -
Does the system include human checkpoints?
Model routing is not enough when the next action can change real systems.
These questions sound less exciting than benchmark charts.
They are also closer to how AI becomes useful inside a company.
How this connects to Buda
Buda is built around the idea that humans manage agents, not the other way around.
That matters more when teams use multiple models. A useful AI agent workspace needs more than a model picker. It needs context, execution, permissions, review, logs, and a way to decide when a task should escalate.
In Buda, teams can organize knowledge in Drive, run agents in sandboxed workspaces, connect workflows through channels and skills, and keep humans in the review loop before important work ships.
The goal is not to use the cheapest model every time.
The goal is to use the right model for the right step, with the right human judgment around it.
That is where cheaper models become powerful: not as a compromise, but as part of a managed agent system.
Explore agent workflows in the Buda dashboard.