Stick with Cursor or switch to Claude Code? 🤔

Discussions
22

None of the three is “perfect”

Three? Where did that come from?

There are two different versions of Claude Code, the terminal app, and the desktop app, and they have subtly different characteristics.

Claude Code (both versions) burns through tokens much faster than Cursor does when you’re being interactive.

This is because Claude Code tends to dump more of the entire referenced file tree into the LLM (actually, each subagent is trying to do its own analysis of what it needs), wheras Cursor has a whole bunch of clever technology (RAGgish) to only send up the code it actually needs to. Whatever the internals, Cursor (interactively) seems to use far fewer tokens.

The above, alas, isn’t true for using Cursor more agenticly. If you want to do this, Claude may well have an advantage.

Claude Desktop, on the other hand, it easier to make run in “full auto” mode from a GUI.

In Claude you can set up CoWork with a prompt like

Review the Trello of ‘to do’, and pick three tickets that you could work on next. Give me a recommendation as which you would work on, with reasons in support of all three. Wait for me to approve your choice before starting…

… and it will toddle off. If it’s possible to do this in Cursor, I’ve not worked out how to.

Claude CLI is easier to automate into something you just trigger via shell scripts. You can also do this with Cursor with the --terminal-query “Your prompt here” option, but I’ve not tried that so can’t comment on how reliable it is

Cursor is much more reliable when it comes to uptime of the backend LLM.

It really is. Claude is, in April 2026, about as reliable as the electrics on a 1980s British Sports. (Insert joke about how Lucas was denied the patent on “darkness” but awarded the one on “sudden, unexpected, darkness..)

Uptime is a feature. Just because it’s (hopefully) an invisible feature doesn’t mean you should ignore it.

Claude, to be fair, is better if you are a European who wakes up early, but by the time most North Americans are awake, its server farms are more overloaded than an IKEA bookcase trying to carry the contents of a fully loaded Kindle.

But, provided you are mostly in Cursor’s Auto mode, Cursor’s uptime is much better. (Obviously, if Anthropic fails, then it fails for everyone, including those of us using it through Cursor’s IDE).

So, how to use Cursor for cost optimisation?

Always start Cursor in Plan mode. See below for which model to choose.

Allocate about 75% of your personal time to reviewing the plan in detail.

Once you have approved the plan, switch back to “Auto” mode, so Cursor can use a much cheaper model. You get much more “Auto” usage than if you choose the model by hand, and Cursor can act as a “buffer for inference provider downtime” by switching to an “actually available right now” mode.

Choosing the right model for the plan involves remembering that planning is input heavy and execution is output heavy, and higher-end models are much more cost effective at input heavy tasks.

  • If your time is expensive (say, you’re paying your people Silicon Valley coder salaries) choose the best model your budget allows.

  • If your time is not expensive (say, a hobby project), start planning in Auto, and re-do the plan from scratch in a more expensive model only if you need to.