If your coding agent costs more than $2 per feature, read this. You're probably paying for 3 things you don't need: Opus reading every file in the repo. Opus making choices a $0.042 model can make. Simple tasks running on the most expensive model. Fix those 3 and the... same agent drops from $6.42 to $2.00 per task. Full architecture with code below 👇show more

Bootcamp prompt hoarders celebrate 80% API bill reduction by stacking hybrid multi-model wrappers thinking they cracked the autonomous agent architecture. They stare at temporary token price deflation while remaining completely blind to the catastrophic cascading runtime latency tax hiding inside their nested orchestration loops. Shallow financial optimization is a dangerous proxy for real performance engineering. The moment your non-deterministic pipeline hits real-world concurrent session traffic, your lightweight routing gates choke on implicit output drift. Pulling un-profiled multi-tier models to bypass frontier token rates doesn't fix architectural fragility; it just route-forwards your processing debt and leaves your distributed backend infrastructure completely I/O blocked.

good share buddy

yeah the repo re-read is brutal, been there man

seems like dialing things back a bit could save a ton. fr keeping it simple is underrated 🙏

smarter model routing can save serious money without sacrificing quality

Routing is the biggest lever, agreed. Two things to add from running agents in loops: a hard retry limit per step, and an escalation path so a weak answer from the cheap model goes to the strong one. Without the limit, the savings disappear when one step keeps retrying.

the opus reading every file one got me lol

Great read!

The $2/feature tax is usually three Opus habits: full-repo reads, enum picks, and trivial tasks. Default those to a cheap model; keep Opus for novel design. Routing policy is the product when overseas margins can't eat frontier tokens on every click.

charcloth trick changed my firekit forever

Most coding agents waste money because they process more context than the task actually needs. Giving the right model the right context matters more than using the most expensive one. That's something we're thinking about at @jarvixdotlive too.

routing the cheap yes/no checks off opus is the real cut. we waste so many frontier tokens on "is it done" loops that a tiny model could own

the repo re-read is a context problem before it's a routing problem: the agent walks the tree because nothing durable tells it where things live. the biggest cut for us was a maintained project map, so it opens the three files that matter instead of re-walking the tree.

Sounds like my code’s on a luxury diet, time to cut the caviar and stick to ramen.

the first one hurt me the most. I once watched Opus re-read my entire repo three times for a one-file change, and I could feel the bill ticking. splitting planner and executor onto cheaper models cut my costs to pocket change.