Case study · August 1, 2026
The night my AI agent finally felt like a coworker
Handed twenty products to an agent before bed, woke up to finished marketing copy at 1.6 cents total. What actually changed once Reality Router started picking the models.
I dropped a CSV of twenty tea and coffee products into a folder on my Jetson before bed, told an agent to write a marketing description for each one, and went to sleep.
In the morning: descriptions.csv was there. Twenty rows. Each product covered, each description in the right voice, each within the length range I'd specified. I spot-checked half a dozen — they were fine. I would have written them the same way.
The receipt: 1.6 cents.
That's not the point.
The point is that I actually believed the agent would do it. I didn't set a timer on my phone to check whether the loop had gone off the rails. I didn't wake up wondering if my credit card had a $47 charge on it from a runaway iteration. I didn't have to babysit.
Why this time was different
Every previous time I've tried to hand real work to an AI agent, some fraction of me stayed anxious. I knew I'd left it with an expensive model and a task that could balloon. I knew each retry cost money. I knew that if it got confused, it would keep going, cheerfully burning tokens on the wrong thing.
This time the router picked the model per prompt. Twenty short generations, all handled by a lightweight model that costs less than a fifth of a cent per description. If any single generation had needed a stronger model, the router would have reached for one — but it didn't need to. Twenty short marketing blurbs are exactly what a cheap fast model handles well.
The bill couldn't have blown up because the router wouldn't have let it.
What the receipt actually looks like
| Products described | 20 |
| Total agent time | 4 minutes 16 seconds |
| Total cost | $0.016 |
| Cost if I'd left it hardcoded to Claude Opus 5 | $0.63 |
Would I run this workflow every week to keep my catalog fresh? Sure — that's an 80-cent-a-year subscription for a task that used to sit in someone's job description.
Would I hand it something bigger — 200 products, 500 products, a whole seasonal refresh? Yes. Same shape. Same lightweight per-item calls. Same predictable receipt.
What actually changed
Not the models. Not the agent. Not the task.
The router. Reality Router in front of my usual OpenCode setup means every prompt goes to the cheapest model that can actually do it. I stopped picking models. I set a preference for cost. The router handles the rest.
That's the whole shift. My workflow doesn't feel expensive anymore. So I use it. So I run more things overnight. So the agent starts to feel less like a demo and more like a coworker who works the graveyard shift.
Try it yourself
Reality Router is open source. Point OpenCode (or Aider, or Cursor, or anything OpenAI-compatible) at it, plug in your own API keys, and see if your Tuesday-night backlog gets any smaller.