I was researching and I believe I found the best alternative to Cursor today. I’m going with GitHub Copilot Pro+, which offers 1,500 premium requests for USD 39, plus USD 0.04 per additional premium request. We also have access to many simpler models for free.
It’s important to note that a premium request is essentially a prompt sent to the agent. Regardless of how many tool calls or tokens it uses, it counts as a single request. Depending on the model, it may be charged at 0.33x or up to 3x (for Opus 4.5).
Additionally, with a ChatGPT Plus subscription, you get access to Codex CLI, which is excellent for large tasks. In my opinion, this stack currently offers the best cost–benefit.
For $20 on Cursor, I get about 10 hours of heavy agent use per month (using the cursor Composer 1 model). For $20 on Claude Code, due to their reset every 5 hours, I get 80-90 hours of active use per month (same code base, heavy agent use, Anthropic models). That’s 8x hours (minimum) with Claude Code. Easy decision. Cursor subscription cancelled.
@Guilherme_L I appreciate you sharing your research. I’ve been Claude Coding for several months (other than this recent experiment with Cursor). I’ll try Codex CLI and GitHub Copilot Pro+
I tried github copilot a few months ago. Their harness is not good enough. Same model performed better on other coding agent than copilot. If their harness is good now then its way cheaper than Cursor.
Cursor just passes through API costs. I may be mistaken, but it sounds like you are comparing Cursor’s subscription with additional on demand costs (API billing) with a subscription to OpenAI which does not integrate into the Cursor IDE from my understanding.
Did you install the OpenAI Codex extension in Cursor or in VS Code? If that works for you, then maybe you actually didn’t need the Cursor IDE all along? Glad your figured that out and and stop paying API costs for codex through Cursor, when you could subscribe to OpenAI Codex and use it via extension and get what you needed. Thanks for sharing your experience, but I think you may be comparing different products (Cursor IDE vs CLI), but sounds like your team doesn’t need to work within Cursor. Good luck.
The main issue of Cursor is the Cache Read. Worth nothing cache reads 10x cheaper if Cursor uses 10x more cached tokens.
Simple tasks using 1-2M tokens? That’s nonsense. (And no, it’s not because MCP, or whatever. I disabled all of them)
I don’t know why they started to send much more data as cached data to the LLM providers. Price is the same, but Cursor started to spend a huge amount of unnecessary Cache Read. Then, the cost skyrocket.
When using Cursor you need to pay attention at the Cache Read price of the model, not the output tokens price.
Claude Code seems to be much more efficient with context window. Cursor just send everything to the LLM.
I still have my Pro+ subscription, but I’m about to move to CC as soon this billing cycle ends. I love Cursor, its UX, model harness, but the cost simply can’t justify it.
But on general usage over time, they are about equal in cost proportion. So if you know the cost of one (cached tokens), you can calculate the 3X cost of the total query.
I just had the exact same experience. I like Cursor more, but the $100 plan on Codex, while hard to measure, feels like its about 3x the Cursor $200 plan. I went through an entire $200 cursor plan in a week, +$100 credit, +50% model usage (Opus 4.7 and later GPT 5.5).
Its just not sustainable when competitors have cheaper access to the models.
That or Composer 3 needs to be way better than composer 2, which is pretty dumb feeling compared to Opus 4.6/7 and GPT 5.5.
Cursor: I really like your product. Its great. its much better than your competition, and I would pay a premium for it (I have been, for months now). But I can’t be paying $600+ a month when your competition is <= $200 month for the same core service (model access).