Remix clone Hacker News

Aren’t those models 5c per call though? I think that would add up.

Absolutely. I'm currently spending about $50 per workday in additional costs. But it's so much better - an entirely different experience.

▲

vincent_s 7 days ago | parent | next [-]

Max requests are only an additional 5 cents each. The real cost is in tool calls that cost another 5 cents each, which adds up fast in agent mode.

From one day of coding with MAX models:

174 gemini-2.5-pro-exp-max requests × 5¢ = $8.70

1269 premium tool calls × 5¢ = $63.45

143 claude-3.7-sonnet-thinking-max requests × 5¢ = $7.15

▲

bn-l 7 days ago | parent | prev [-]

You’re spending $1500 in additional costs? How?!!? I can’t even conceive of how I would spend that much with cursor. What am I missing? Are you ultra productive or just inefficient with tokens?

▲

vincent_s 7 days ago | parent [-]

Being inefficient with tokens actually makes you super productive. It's too expensive in the long run though.

The last few weeks have been quite frustrating with Cursor. I dove deep into the issue and figured that the most annoying problem - which leads to all those frustratingly poor replies from the LLM - is how Cursor cuts down the context. You can test this yourself: just add a long file to the chat and ask if it can see the file.

Recently I discovered that all these problems disappear with the "max" models. This is exactly what I wanted. The price of 5¢ per request is manageable, the real issue is the cost for tool use in agent mode (see my other comment).

▲

bn-l 7 days ago | parent [-]

thanks for the reply. Do you have a write up on how you use cursor?

	▲	vincent_s 6 days ago \| parent [-]
		No write-up yet - Cursor iterates so fast that any guide would be outdated in a few weeks. My tips: - Check out the Cursor docs. They're concise - read through them to understand the features and model/context behavior - It's basically all chat now. Chat has manual mode (previously edit mode), ask mode, and agent mode - For one-off file changes, use manual mode. Just tell it what to do, and it shows changes as diffs you can accept/reject - Agent mode is similar but the model can use tools to read files not in context (plus some other stuff like run commands and search through files). It works in a loop until the task is complete - reading files, editing them, fixing lint errors, reading more files, etc. - For agent mode, Claude Sonnet works best. Other models sometimes fail to use tools correctly or just stop mid-task - Context is critical. Works best when you provide all necessary context, or at least file names/directory trees in agent mode - Biggest issue is context cutting. Cursor truncates files and doesn't give the LLM all the code you think it does. Even in max mode, the read file tool only ingests up to 750 lines per file (though I think actively adding files to context lets it read more in max mode). Sometimes copy & paste the file contents into the chat prevents truncating. - This is why I use max mode for almost anything beyond simple small file edits