“Now I fundamentally think of cursor as two interfaces: the thing that organizes my projects and cloud agents, and the thing I'm focused on now (IDE). Love em, hate em, dont understand em...”
“Hey, I've been using Cursor desktop for the past two years, mainly with Claude Sonnet 4.x models for all work. I'm trying Cursor Agents (cloud) now so it works on the go, and tried some planning work with Claude Opus 5 (High) and some coding work with Codex 5.3 (High).”
“Cursor does this because it is repositioning itself from an AI-enhanced code editor into an agent orchestration product. Making Agent View the default drives users toward that new workflow, even when it damages the familiar IDE experience.”
“It’s still buggy, every patch that comes out I hope polishes the quirks more (terminals disappearing / closing even when running while you’re in other agents, starting new sessions from a project starts in the wrong workspace silently, notifications constantly getting stuck as unread, etc) - but I’ve really adapted to the flow of it.”
“cursor does break files into retrieval chunks, and that does cut tokens per call. the catch i hit is the fragment set changes every turn, so the prefix never repeats and cache-read goes basically to zero, while the whole-file path keeps reusing the same prefix at cache-read price.”
“same model, different harness: cursor sent about 3x more tool calls per task for me because it kept re-reading files claude code had already cached. what fixed the token bill was a .cursor/rules file capping file reads and forcing a plan first, went from \~90k tokens on a refactor to \~35k.”
“I have cancelled cursor after it was bought but I preferred it because I felt I had more of an idea of what was going on and it felt more responsive for small changes.”
“Honestly it's the tab completion and how it grabs context from multiple files at once. codex feels like it only really gets one file at a time, and claude's smarter when you need real reasoning but it's not built into the editor the same way. if cursor ever falls behind on that autocomplete stuff i'd switch in a second, i'm not attached to it.”
“Use Cursor if you want to use more models. The is no point going to Cursor only to use Opus or Fable - they will charge full api price as premium models. And Fable can burn $100 usage in matter of minutes depending what you do.”
“I have it linked to GitHub and use it as the orchestration layer to all the development agents in Cursor. I use High when needed, but what I don’t do is use ChatGPT “Work” over “Chat” - that’s when it drinks the usage, in “Work”.”
“for a long time cursor was ahead of most if not all. it's why i used it for the past year and a half. now everything has caught up and i think the only remaining truly unique benefits are super low cost grok and the ios app?”
“Yes, Cursor is expensive as you have to pay direct API rates. Grok 4.6(which is better than Luna) and Composer are cheap in Cursor and have a large pool.”
“Auto works well for generic web/soft dev, but for gamedev, specifically when it comes to visual placement of 2D components in a 2D UI, it kinda sucks.”
“It’s been a while since I used a ChatGPT sub but I easily hit the limit within the 5hr window. That’s never happened for me on Cursor (with first party models)”
“Cursor is great for throughput and background tasks, but the real cost shows up after the MVP: reviewing architecture, adding tests, checking security, and untangling generated code. For anything users depend on, I’d keep an explicit review and hardening pass rather than treating “it runs” as shipped.”
“claude is great for logic, but cursor sees my entire codebase. with cursor, it feels less like chatting with AI and more like pair-programming with someone who knows my project”
“Looking forward if Grok 4.7 is Sol-esq, and there is a composer 2.5 replacement like Luna, and Grok 4.6 stays decent for the cost, the cursor subscription will basically be identical on value to Codex with much better cloud agents and better multi-api support.”
“What Cursor has is millions of people with it installed and a decade of habit around an editor. Worth a lot of money, and not the same thing as being the best tool.”