zed
zed copied to clipboard
zed-industries
Agent usage calculation
Summary
Is counting usage based on per api call? compared to competition where it's based on per prompt.
I've had the agent trial access for about 15 minutes, started three threads with the agent which achieved a skeleton react/tauri app (with 1 button) - that doesn't even actually start, and blew through literally all of the trial usage - even though the trial usage is supposed to provide 150 prompts.
I think this would make Zed about 50x more expensive than cursor? seems like maybe a bug considering the api usage is based on tokens not per api call.
Zed Version and System Specs
Zed: v0.185.9 (Zed) OS: macOS 15.4.1 Memory: 16 GiB Architecture: aarch64
Yeah even with max mode off the agent still uses additional prompts itself which doesn’t align with what I’m reading on the blog post
imo, the usage calculation is not very transparent or something is wrong! I burned through my pro credit within a week without even using it heavily.
Just enabled "Usage-Based Spending" and use 1 Prompt with "Max Mode" on. The Agent changed 3 files and it costs me $1.15. How is this possible, I don't understand 🤷
In my experience the Zed Pro pricing is indeed per prompt, as long as Max Mode/Burn Mode is disabled. When using agentic editing, after 25 requests a new prompt is required to continue.
GitHub Copilot calculates usage per request, and every tool call counts as a request. When their premium requests billing starts this means you'll reach the plan limit in a few prompts.
GitHub Copilot calculates usage per request, and every tool call counts as a request. When their premium requests billing starts this means you'll reach the plan limit in a few prompts.
Just a note, it doesn't seem to be the case in vscode according to what I've seen, see also evidence in https://github.com/zed-industries/zed/discussions/31068.
For me, this is one of the biggest limitations when it comes to using Zed at work. We have a GitHub Copilot Enterprise license, but with VSCode it seems that "premium requests" are consumed more prudently. Maybe now that the Copilot Chat code has been open-sourced, it will be easier to investigate what's going on.
This issue is related to Zed's hosted models. Discussion about Copilot should occur elsewhere, such as in the discussion mentioned above.
There is now much better documentation describing how usage is calculated here:
- https://zed.dev/docs/ai/plans-and-usage
In the 2+ months since this issue was opened we have also fixed some bugs. If you are seeing issues with prompt accounting, please open a clean issue with explicit details with how to reproduce or if you believe you have been incorrectly charged you can reach out to [email protected].
Thanks!
