@tenobrus Just had a session with 8 turns and 270K context, 80M total tokens processed turns != API calls though this isn't it
an average claude code session has just *four* turns
Each turn outputs an average of 2,400 words.
@tenobrus Just had a session with 8 turns and 270K context, 80M total tokens processed turns != API calls though this isn't it
an average claude code session has just *four* turns
Many users reacted negatively to Claude Code sessions averaging just four turns, calling them disappointingly short and too unproductive for meaningful conversations.
No Digg Deeper questions have been answered for this story yet.

checking in on my own stats over the last few months:
median: 7 mean: 30 p95: 59 max: 1271

Most of my projects are done in the same continuous multi month bagillion compactions session that only closes when my computer crashes, lmao, even then I use resume. That's because I keep making things with an unfinished pipeline that I am continuously actively working on with Claude. Using the same session gives him context on the latest stuff we are working on that may be half implemented or such but still partially usable.

@thinkingavocet @tenobrus https://www.youtube.com/watch?v=LIaJ0huGoZk

this is why I strongly suspect the claude subscription plans are actually profitable for anthropic in aggregate gross profit. the barely using it users are massively subsiding the power users. the subscriptions probably don't pay for the training runs and all the r&d though, unlike the api rates, which is I suspect the main the reason for anthropic's crackdown on purely programmatic use of subscription plans: it's a danger for their economic model

@tenobrus ????

@tenobrus “Typical” sounds more like it means median, not average

@tenobrus i had median: 16 mean: 28.3 p95: 89 max: 393
(if you exclude sessions where i didn't send any messages at all, which actually excludes like a majority of the sessions)

@tenobrus I don't think I've ever had a Claude Code session under 100 turns 😂

@tenobrus My median session is definitely 1 turn.

@tenobrus that's me loading up an agentic harness to batch rename files
@tenobrus reward hacked
an average claude code session has just *four* turns

@tenobrus Sounds right. My stats over the past month, measured by count of UserPromptSubmit hooks:
user turns: sessions count with that many turns 1: 124 2: 76 3: 38 4: 34 5: 19 6: 12 7: 7 8: 5 9: 3 10-14: 8 15-19: 2 20-29: 1 30-49: 1 50+: 1

@henrytdowling @tenobrus Gonna channel @besttrousers here and say it's both. It's definitely a median, not a mean, but a median is a kind of average.

@tenobrus thats because you gotta start a new session often

@ja3k_ @tenobrus why?

@henrytdowling @tenobrus @besttrousers Although as written I suppose they could be talking about the mode

@i2equalsj2 @tenobrus What makes it last that long? I’m curious how one session could have so many turns

@jltvar oh yeah 100% this is very clear

@tenobrus Where can you see this?

@henrytdowling @tenobrus @besttrousers Anyway, colloquially an average can refer to any of those, and median is quite often the best option