Opus consumes a lot of tokens when taking / analyzing screenshots

Support Help
,
1

I’m checking https://cursor.com/dashboard?tab=billing regularly to monitor my token usage and I noticed that when the Opus agent analyzes images as part of its work (sometimes it also captures screenshots of the built-in browser tab on its own…), the token usage / dollar about increases very, very quickly.

Usually, when doing only coding / text edits, I am used to seeing a usage that is something like a couple of cents per minute, but when it starts working with image assets / screenshots, it can easily be a dollar per minute.

I had hard time finding how this is supposed to work, or if it’s model-specific (Opus seems to consume especially a lot but I didn’t do a proper comparison with other models). Is this high usage expected?

3

Yes, this is expected behavior. Images are tokenized by vision-capable models, and a single screenshot can easily consume a lot of tokens - Claude Opus is also one of the more expensive models, which compounds the cost. If you want to reduce usage, you can switch to a cheaper vision-capable model (like Sonnet), disable the browser tool, or avoid including images/screenshots when they’re not essential.

1 Like
4

Hey, thanks for the request.

Yes, that’s expected behavior. Vision or image tokens are a lot more expensive than text tokens across providers. When the agent analyzes screenshots or images, usage can jump by several times, especially on Opus, which is already one of the most expensive models.

A few ways to optimize:

  1. Use cheaper models for image analysis:
  • Sonnet 4.5 or Gemini 3 Flash can handle vision tasks much more cheaply
  • Save Opus for harder coding tasks that don’t involve images
  1. Limit how often screenshots are taken:
  • You can set up tool approval in Agent Settings to control when the agent takes screenshots
  • Manual approval mode lets you see each screenshot request before it runs
  1. Monitor the usage breakdown:
  • In the dashboard you can see a detailed breakdown by tokens and models
  • This helps you spot where most of the spend is coming from

If you want, share a few examples from your dashboard usage breakdown and I can take a look at the numbers.

1 Like