Every company running agents in production has the same chart somewhere: token usage climbing exponentially, and a finance team asking which line goes up next. Most answers involve caps, alerts, and asking engineers to use the expensive model less.
Coinbase's answer involved none of those, and it cut the bill roughly in half while usage kept growing. It is the most concrete public playbook yet for what open weights are for on an ordinary day, when nothing is on fire and the only question is what a completed task should cost.
Half the Bill, No Caps
In late June, Brian Armstrong posted a thread on how Coinbase manages AI spend for its engineers. Not by rationing. The opposite:
Armstrong
"How to keep AI spend flat while token usage grows exponentially: Not with friction and spend alerts. With better defaults, routing, and caching."
The mechanics are three moves, and note that every one of them depends on open weights existing:
- Better defaults, not usage caps. Engineers can pick any model they want, but the default is a cheap open-weight model, GLM 5.2 or Kimi 2.7. Frontier models by choice, not by habit.
- Prompt-aware routing. Requests get scored and routed to the cheapest model that clears the quality bar for that task, instead of everything going to the most expensive endpoint by reflex.
- Cache-aware requests. Restructuring prompts so caches actually hit. Coinbase took the cache hit rate in LibreChat from 5% to 60%.
Armstrong's framing line is the one to steal for your own board deck: "The goal isn't fewer tokens used, it's fewer tokens wasted." The result was AI spend cut roughly in half while token usage kept growing. Same tokens, half the bill.
Gergely Orosz surfaced the thread to the wider engineering world and asked the right question:
Orosz
"Coinbase seems to have lowered their token spend ($$) to about half, by 1) routing to cheap inference like GLM 5.2 and Kimi 2.7 that are still pretty performant 2) Smart routing + caching. They still use the same tokens as before. Start of a trend?"
The replies carried the obvious objection: forcing engineers onto weaker models is "the equivalent of providing people a Chromebook instead of a MacBook Pro." The counter landed just as fast: "GLM-5.2 is not less powerful tho?" That exchange is the whole debate in two lines. The objection assumes open weights are the budget tier. On a growing share of real workloads they are simply the correct tier, and a router that knows the difference is worth more than a discount. That routing thesis is the subject of Token Prices Fell. Bills Went Up., and Coinbase is what it looks like executed.
Set this beside the price collapse and the mechanism is complete: the floor of capable models dropped 99%, and Coinbase is what it looks like to actually stand on the floor. Owned defaults absorb the volume. The frontier stays available for the calls that deserve it. Nobody files a ticket to get their model back.
The control argument for open weights lives in Part 4, and it is about the worst day of your year. This one is about every other day.
Route it. Own the defaults.
This is Part 5 of the Open Weights series. New here? Start with Part 1, What Are Open Weights, Actually?
