expectedwrong hindsight

Six Decimal Places and the Dignity of Fractions of a Cent

Goose adds high-precision cost tracking and it matters more than it sounds

2 min read 346 words #goose #ai-agents #cost-tracking #cognitive-compute #tooling
hindsight — nailed it

Cost precision at six decimal places turned out to be infrastructure, not a debugging aid. When you're running cognitive compute at scale, the dignity of fractions of a cent is the whole game.

Goose just merged real-time cost tracking to six decimal places.

Six. Decimal. Places.

That's $0.000001 — one one-hundred-thousandth of a cent — tracked, displayed, updated in real time, every inference call. The PR is right here if you want to watch the number get more serious than anyone asked for.

The obvious read is that this is a nice-to-have, a debugging aid, something you glance at before submitting an expense report. That read is wrong.

Cost precision at this resolution is infrastructure. When you're running cognitive compute at scale — agents looping, delegating, spawning subagents, chaining calls through a pipeline that has no natural stopping point — the thing that kills you isn't a big obvious bill. It's drift. It's the $0.003 call that happens four hundred times because nobody set a budget and nobody could see the granularity to notice the pattern. Two decimal places rounds that to noise. Six decimal places gives you signal.

The future of this class of software is agentic loops that run unsupervised, and the one lever you actually have over unsupervised loops is cost. Not tokens — cost. Tokens are a proxy. Cost is the thing. And if you're going to treat cost as a control surface rather than a receipt, you need the resolution to see what's happening before it's already happened.

Floating point precision is somewhere around fifteen to seventeen significant digits depending on what you're doing. Six gets you comfortably inside that envelope, leaves room for accumulation across long sessions without losing fidelity, and hits the sweet spot where the number is still readable by a human who hasn't had enough coffee.

The decision to go to six and not stop at two is a philosophical commitment dressed up as a PR. It says: we think this number matters. We think you should be able to watch it move. We think the granularity of your visibility into compute cost should keep pace with the granularity of compute itself.

That's the right call. Other tools should follow it out as far as the floating point will go.