Show HN: Claude's innerworkings: turn 170 costs 2.1x turn 20, over 14,640 turns
On Aug 26, this happened in my CoWork session:
Tom: I can't help but think we can do some other cool stuff with this knowledge of Claude's innerworkings regarding sessions, etc. Any thing sticks out to you that I haven't captured/asked for yet?
Claude: Six: the scaffolding trend (the cost of our own rules, creeping 87,999 → 92,619 → 93,809 across three seats), the review share of the wall clock (the thesis measured), the session aging curve (when a fresh seat is cheaper), before/after receipts on process changes, model economics, and a crash watchdog reading both records.
Four days later, we analyzed 144 sessions, 14,640 turns. The facts showed us that past turn 141, a turn costs 2.1x what a turn cost in the same session's first twenty. https://my-status.app/stats#aging The worst offenders were our own batch sessions, so we stopped batching. Wanna see your own curve? The script is public: https://github.com/lordbron/mystatus-samples/tree/main/claud...
What we do is our own flavor of "one thinking session, short task sessions" to better utilize my plan. I kept running out of Fable usage and had about half my usage leftover for Opus.
But this only happened because we wrote a script to surface the data to us. (Granted, our first run found a bug in our script: 46.6M tokens that were really 28.1M. https://my-status.app/t/t33hPWp7zmtnAE5G/u/d2ba812d-13bb-4ad...)
This peek into the innerworkings of Claude fascinated me, so we kept counting: the standing context re-read every turn (114,133 tokens on the hub seat https://my-status.app/t/6iYmualpgrZ6_C9u/u/2f4d4498-dcdb-43a..., 108,213 on a spoke https://my-status.app/t/Zsse45gEDBTQ6mOL/u/c3f8bca2-b7f9-4be...) and the human's share of the clock (67% https://my-status.app/stats#review).
The best part of all this to me: every feature moving forward comes with receipts. Take the aging curve: it cost 2.6 person hours and $8.23 in tokens at list rates. https://my-status.app/receipts
We have a whole bunch of stats for you to nerd out on if you want. It's a bit overboard, but new things pop up as other things come down. https://my-status.app/stats
I know we're not the only ones geeking out about stuff like this. What else have you all measured, and what surprised you in the findings?
Oh and if you wanna know what the heck the project was originally built for: it's public task tracking, for things like a 20 mile run https://my-status.app/t/EdiPcBGBZxcGkr9x or which app versions are supported https://my-status.app/t/i3jQU03T46e0hrsk. The experiment behind the project is: Can a human and Claude build something new, splitting the work along what each excels in?
Interesting!
Is there any loss of momentum in starting over vs. eating the extra cost? My (admittedly limited) understanding is that you build on the previous interactions, the model learning to some degree what you want, how you want it, etc.
Do you have to start from zero each time?
No, let me expand using some of the items in my post.
When I talk about my own flavor of "one thinking session, short task sessions", it takes into account exactly what you're asking. The thinking sessions write to a series of documents, throughout the session. So if I'm talking about how I want Claude to code or how I want it to think about my voice, it takes notes in these documents. Meaning, when I shut down that thinking session, it dies and all the in-memory context dies, BUT the important stuff it learned got saved to the docs. Those documents become the scaffolding of newer sessions. When I spin up a new thinking session, Claude reads that scaffolding first, then says "Okay, I know everything about you and your project. We left off with these three things, which do you want to work on today?" The chart on our stats page actually prints that trade: at 2.1x, a re-seat that costs ten turns of getting up to speed pays for itself after ten more turns of work. Past 180 turns, at 2.86x, after six. https://my-status.app/stats#aging
It's that same scaffolding that we saw creeping up over time in this bit: "the scaffolding trend (the cost of our own rules, creeping 87,999 → 92,619 → 93,809 across three seats)" That meant that over three sessions, my initial token count for "getting up to speed" increased by roughly 6K tokens. "Don't let me repeat myself." and "Don't try to make a new table for everything." and other such rules got added during those sessions and thus upped the token size of the scaffolding. But that was okay to me because it meant there was no penalty for me killing sessions.
Now, what's interesting is a couple of days ago Anthropic cut cache-read pricing by another 75% on Fable 5.1. That was sweet music to my ears, because they basically said, "The way you work is smart and we should reward that." LOL That's not what they really said, but that's how I read them. :)