Has anyone else seen a huge increase in Agent token usage since August 1?

Ideas Feedback
1

I’ve been using Cursor for over a year and have been working on a fairly large multi-app project for the last couple of months. Until recently, I was very happy with it and regularly recommended it to other people.

Has anyone else noticed a dramatic increase in agent usage this month?

Previously, I could keep a thread open and work on substantial application infrastructure without my usage skyrocketing. My normal Pro allowance generally lasted the entire month.

Since August 1, the difference has been extreme.

I’m now regularly seeing ~1M tokens consumed for relatively ordinary agent responses, including in fresh chats. A normal task/session can consume 2M+ tokens, and I’ve seen individual totals reach 17M+.

What concerns me more than the raw token count is how those tokens are being used. The agent frequently seems to:

  • read files

  • reread the same files

  • reason about what it needs to do

  • inspect things again

  • make limited progress

  • then ask whether it should actually perform the task

I’ve seen this happen even with relatively small requests such as hooking up an existing button.

On August 1 alone, I exhausted the Pro usage that had previously lasted me roughly an entire month, despite not doing unusually heavy development work. I upgraded to Pro+, but I’m already at 32% of the included Cursor-model usage, while 100% of my additional model usage has gone toward other models.

I normally used Auto, so initially I wondered whether this was a routing issue. However, I’ve tried lower-cost models such as Composer and am still seeing unexpectedly high usage.

I’ve also tried adding stricter Cursor rules to reduce repeated exploration and encourage the agent to execute tasks directly, but that hasn’t made a substantial difference.

The contrast became particularly noticeable after I started using Codex alongside Cursor. On the same project, I’m generally seeing much less repeated reading/reasoning, and it tends to carry a requested implementation through to completion without requiring me to start a new reasoning process or explicitly tell it to continue after the initial request.

I’m not expecting a large coding agent to use tiny amounts of context—the project is substantial, and I’m completely comfortable paying for useful compute. My issue is that the increase in usage appears to be accompanied by less task completion, not more.

That’s what makes the current experience difficult to justify. If competing coding agents can complete the same work with less intervention and substantially lower effective usage, it’s becoming difficult for me to continue recommending Cursor the way I used to.

I’m especially disappointed because I purchased an annual subscription based on how solid the product had been for me over the previous year.

Is anyone else seeing a significant change in token consumption or repeated file-reading/analysis behavior since the beginning of August?

5

Thanks for the detailed write up. Most of what shows up in those totals is cached context being replayed on each step of an Agent run rather than brand new input. On a large project, a long session can reach millions of tokens even when the underlying task is small, and cache reads are priced well below fresh input, so the raw token count usually looks more alarming than its actual effect on your allowance. Your usage dashboard breaks each request down by token type, which is the quickest way to see whether a given session was genuinely expensive or mostly cached replay.

The part worth attacking is the behavior you described, where the agent re-reads files and re-reasons before acting, since every extra step replays that whole context. A few things that consistently help: start fresh chats frequently, scope each request to one deliverable and name the specific files involved, and trim any rules set to Always Apply, since those rules and AGENTS.md get injected into every request (rules docs). .cursorignore is also worth setting up to keep generated or archival directories out of context entirely.

6

Hello Kevin,

Thanks for the explanation. I understand how cached context replay can make the raw token total look much larger in a long-running Agent session, and that makes sense.

The issue is that I’m already starting fresh chats extremely frequently—at this point, often essentially every new request—and I’m still seeing the same behavior.

For example, I can start a completely fresh Agent chat with a relatively small, scoped task such as hooking up an existing button. Instead of locating the relevant implementation, making the change, and validating it, the agent may repeatedly read and reread files, reason through the same area multiple times, and consume ~1M tokens or more before completing what should be a fairly contained task. Standard sessions are regularly reaching 2M+, and I’ve seen much higher totals.

That’s why I’m not sure accumulated conversation context explains what I’m seeing. If this were primarily happening after keeping a large thread open for hours, I would completely understand the recommendation to start fresh chats. But fresh chats are exhibiting the behavior from the beginning.

The change is also noticeable relative to my own previous Cursor usage on this same project. I used to keep conversations open considerably longer while doing heavier development work, yet my Pro allowance generally lasted the month. Now I’m creating fresh chats far more aggressively, doing less intensive development, and still consuming substantially more.

I can certainly review .cursorignore and reduce Always Apply rules further, but I’m hesitant to attribute the whole change to project size or configuration because the project was already large before this behavior appeared.

The part I’d really like to understand is whether something changed recently in Agent’s exploration/tool-use behavior or Auto routing that could cause substantially more read/search/reasoning steps per task. That’s the difference I’m noticing—not simply a larger token number on the dashboard. Thanks

9

Also curious — has anyone moved to other tools because of this?

Claude Code, Codex, Windsurf, whatever. Does their monthly allowance actually last a normal month for day-to-day agent work, or is it the same mess everywhere right now?

If you switched and it felt better, what did you switch to?