why long ai chats slow down (and cost more).
you keep one chat open all day, and by afternoon it is slower, foggier, and quietly more expensive. that is not your imagination, and it is not the model getting tired.
a long ai chat slows down because the model re-reads the entire conversation every time it replies. each new message adds to the pile, so every answer processes more text than the last. speed drops, usage climbs, and the cost of the whole session grows faster than its length. the fix is not a better prompt. it is a fresh session.
here is the receipt that this is real, not a feeling. in the week of july 20, 2026, anthropic shipped an update to claude code, its coding tool, that fixed a bug where long sessions slowed down quadratically (release v2.1.216, with v2.1.218 following on july 22). the people who build the ai had to hunt down session-length cost blowup inside their own product. if it bites them, the all-day chat on your screen is a milder version of the same thing.
why every reply gets a little slower
each turn, the model reads the whole thread from the top, not just your latest message. it does not keep a running memory of the conversation the way you do. the transcript itself is its memory, so to answer message two hundred it re-reads messages one through one hundred ninety-nine first. a three-message chat re-reads three messages. a three-hundred-message chat re-reads all three hundred. that growing pile of reprocessing is what you pay for, in seconds and in usage.
why the cost grows faster than the chat
the total work of a session grows with the square of its length, not in a straight line. a chat twice as long is not twice the work. it is closer to four times, because each of those extra turns re-reads a longer and longer history. that is what "quadratic" means, and it is why the last hour of a marathon thread costs far more than the first. anthropic's july fix was about a version of this getting worse than it needed to be. the underlying shape, cost climbing faster than length, is just how a chat works.
speed and cost are only half of it. a long thread also gets less accurate, because the model weights the start and the end of its window and lets the middle fade. i wrote up that failure mode in why pasting in everything makes ai worse. so a marathon chat is the worst of both: it costs the most and it is where your earlier instructions are most likely to get lost.
one long thread vs fresh scoped sessions
the answer is not to chat less. it is to chat in scopes. here is the same work, done two ways.
| what you notice | one long thread | fresh scoped sessions |
|---|---|---|
| what the ai re-reads each turn | the entire day so far | just the task in front of it |
| speed as it grows | slower every hour | fast, resets each session |
| usage and cost | climbs with the square of length | flat and predictable |
| answer quality | the middle of the thread gets lost | the task stays front and center |
| best for | one continuous task you are actively shaping | switching between different jobs |
when to start a new chat
start a fresh session whenever you switch tasks, and whenever a thread gets long enough that replies feel sluggish. four rules that keep sessions cheap and sharp:
- new task, new chat. drafting an email and then analyzing a spreadsheet are two jobs. do not make the model drag the first one through the second.
- carry a summary, not the whole history. when you do need continuity, paste a short recap, or better, keep the context in a file the model can read. that is the whole idea behind a files, not chats context system.
- put your rules in a fixed place, not in the scroll. a standing instruction the model reads at the start of every session beats re-explaining yourself in message two hundred. that is what a skill file is for.
- watch for the fog. when it starts forgetting what you said an hour ago, that is the middle-of-the-thread problem, covered in why chatgpt forgets everything. the answer is to start over, not to paste even more.
the takeaway
you do not need to use ai less. you need to use it in scopes. one long chat feels efficient because everything sits in one place, but it is the most expensive and least accurate way to work: slow to answer, easy to lose the middle, and climbing in cost the longer it runs. short sessions, a summary to carry forward, and your rules in a file get you the same output, faster, cheaper, and sharper.
want your context to survive the reset?
the ai builder toolkit is a set of claude skills that hold your rules, your voice, and your process in a fixed place, so every fresh session starts sharp instead of you rebuilding context in a chat that keeps getting slower. ready to install and run in claude today. no course, no call.
see the toolkitor just follow along. new field notes most weeks on x, instagram, and tiktok.