Nobody picks models by hand
Choosing a model per subtask, every time, is tedious. So nobody does it.
Horsy sends each subtask your agent spawns to the cheapest model that can do it, checks the result, and moves up one model if the check fails.
Then /horsy audit shows what your last sessions would have cost. Changes nothing. Not published yet.
cheapest first → verify → step up only if it fails
Illustrative demo. Numbers show the format, they are not measured.
The problem
The agent reports success. You find out later. The rework costs more than the original task.
Renaming a variable, grepping call sites, writing a test stub. All on the most expensive model, burning quota or money.
Choosing a model per subtask, every time, is tedious. So nobody does it.
Every parallel agent and every CI run multiplies the waste.
vs Opus-only, on a replay of 200 real agent steps from an outbound sales workflow, at 95% agreement. Not yet tested on coding tasks.
planted errors caught by Horsy’s check, against 4 for a small model used as a judge. $0.01 vs $1.10.
How it works
That loop never changes. What changes is how much Horsy is allowed to do, and it earns each step on your own repo.
Every “done” gets checked before it is accepted. If the work isn’t there, you hear it now, not in code review.
Horsy watches your sessions and tells you what it would have routed, and what that would have saved. Your models stay as they are.
Turns on only for task types where your repo’s data shows it costs less and breaks nothing. Everything else stays on your default.
The receipt
Model, cost, time, pass or fail. When a step cost more than running Opus alone would have, the receipt says so.
Guarantees
These are product rules, not marketing. If it doesn’t pay off, it turns itself off.
/horsy off, pin a model per task, one-line receipts.Placeholder · beta stats go here once they are real
Built in public. Every number on this page will come from receipts you can inspect. Follow along on GitHub.
For teams
Same quality bar, because every routed step is checked. One view across every agent your team runs, locally and in CI, with a budget per person or per pipeline. Your devs keep their workflow. You get the receipts.
Pricing
Two options on the table. Tell us which one you’d rather pay for.
Verify, shadow and routing are free for individuals. Teams pay for the shared view, budgets and caps.
Trade-off: no friction for solo devs, so adoption is easy. But revenue only starts when a team signs up.
Free up to a set amount of routed tokens per month. Above that, a flat monthly fee.
Trade-off: price grows with the value you get. But heavy solo users hit a paywall, and a meter can feel like a tax.