I subscribed to the annual plan last year, and 2025/09/09 will be expired.
However, I’ve noticed that the plan’s features have been changing frequently recently. For example, the ‘Fast Request’ feature was removed, the amount of free tab-completions is no longer to let people how many they will have, and a new (and expensive) API-based pricing model was introduced.
On the other hand, I’ve looked at other AI coding assistants on the market (like Kiro, Windsurf, Roo with VSCode), and it seems they all have their own problems, based on what users are saying on Reddit and other forums.
So, I wanted to ask for your opinions: Do you think it’s still worth subscribing to the Pro plan annually, or should I just switch to a monthly subscription?
You’re right — after renewing your plan, you’ll probably be disappointed.
If you only have the Pro subscription, it’s really only enough for about 1–2 days of heavy use. On the other hand, it depends on what you actually expect from Cursor.
If all you want is basic code autocompletion, then that’s fine — stick with it. But if you’re used to working with the Auto mode as an agent, then you should definitely renew before September 15, because after that Auto will become paid as well.
If, however, you’ve gotten used to working with an agent backed by stronger models like Claude Sonnet 4, then honestly just cancel your subscription and don’t worry about it. For agent‑style workflows, I’d recommend Claude Code with a yearly plan, combined with Roo Code as the UI shell. From my own experience, this setup is at least as good as Cursor — and in some ways even better, thanks to the built‑in role orchestration.
Yeah, As a researcher / engineer for communication system design, I have considered some solution for AI-aided design (or called vide coding, but I still don’t like this name).
I have used most of LLM model in different platform, including o3, GPT-5, or any well-known.
Your recommendation is very useful and I have taken them into account.
Actually, Tab-completion in Cursor is the most important reason which I still consider to subscribe.
As far as my knows, Other features can be replaced in different way on other IDE or platform but it can’t.
Here are a few recommendations that might help optimize your usage and costs:
It’s been recommended to adjust the rules for how you interact with the AI to keep the responses shorter and more to the point. Additionally, a helpful setting to enable is: Go to: > VS Code Settings → Enable Shadow Workspace
I also recommend using LLM models strategically. For example, use Cursor for core coding and debugging tasks, but for things like .md file creation, git questions, hosting questions, or other general requests, consider using a different AI provider that might offer free requests.
You’ll also have a better experience if you make small manual changes yourself. Don’t use the AI for every minor alteration, like changing “Hello Peter how are you” to “Hello Peter, how are you?”. That’s something you can do quickly on your own.
Furthermore, you have the option to modulate or modify an agent’s behavior with a specific prompt to minimize the output as much as possible. For example:
Your primary goal is to transform my instructions—no matter how simple—into precise, context-aware, and professional prompts before you execute the actual task. This transformation process is designed to ensure clarity, minimize errors, and deliver the highest quality results.
I would also advise moving away from the “Thinking model” and “MAX” modes when you are using $20 monthly instead of Pro + which gives you more room to use high-end llm optimized features within Cursor feature-lineup.
Instead, try breaking your work down into smaller to-dos that can likely be handled sufficiently by a model like Claude Sonnet 4.
I’m not sure whether you Cursor Staff know until now,
But this attitude is the main I consider if I need to quit.
I trust that most of people also think about that.
Perhaps the thing I will said is a little rude,
but…. You’re talking as if I haven’t done all that before.
As a Cursor Ambassador who has used the IDE for over a year, I think the discussion should focus on usage patterns.
For heavy users, the Pro Plus plan is a great deal. I recently spent $25 in a single day using the Anthropic API directly by using Claude Sonnet 4 (NO Thinking).
In that context, the $20/month subscription is far more cost-effective.
The value really depends on your workflow. Let’s have a productive discussion about that.
Committing to anything AI for more than a month is .. unwise at this time. I may even cancel Claude code the way things are developing right now (degrading, Codex gettin better).
Thanks for your rational response. No offence, but:
That argument, "For heavy users, the Pro Plus plan is a great deal...", is a typical straw man fallacy.
API costs are based on token usage. They can also be integrated into any application, which gives users control. Although the length of LLM replies are random, people can use tokenizers to estimate the cost of their inputs and outputs. Cursor doesn’t offer a tool like that.
Furthermore, people like me can’t accept Cursor’s reasoning because their policies lack clarity and honesty, and they have seldom admitted their mistakes.
For example, the pricing for AI models from major companies is always changing. I might argue with those changes, but I can reluctantly accept them because their pricing and how they work are made clear through papers or application notes.
Look at their history: Has Cursor ever done this?
I love this tool, even now. But this erosion of trust is so severe that I have to consider whether I’m out.
Thanks for the honest feedback. We hear your concerns about the recent plan changes, and we’re glad you still find our tab-completion to be the best on the market.
Since you’re exploring alternatives, I’m curious if you’ve tried Grok Code fast (formerly SONIC) and how you feel it compares to Sonnet or GPT?
No, I haven’t tried Grok Code fast, but its benchmark on LLM is not good enough to me (by Livebench.ai benchmark) for writing / programming on signal processing works with formal math expressions.
A news is that I decide to subscribe Cursor again–not annual but monthly.
If I find a better choice, I’ll quit at anytime.
Unless Cursor’s staff finally know that: One lie ruins a thousand truths.
What’s the point? If we want to work with AI, this kind of penny-pinching only slows down the workflow instead of speeding it up. We might as well just open a browser inside the IDE plugin and go directly to ChatGPT or Anthropic. Code autocomplete is only a minor consolation here.
Yes, compared to using the API directly, the subscription has clear advantages.
But when you compare it to competitors, or even to what Cursor itself used to offer, there aren’t any real reasons to stick with Cursor specifically.
Sure, $20 isn’t a big deal if we want to use Agent mode for a whole month. But even at $200 or $400, Cursor won’t give us that freedom—we’d still need to ration usage and worry about running out of credits before the month ends.
No matter how expensive the APIs are for Cursor, that’s ultimately a problem of Cursor’s competitiveness, not its competitors. I genuinely want to know what Cursor’s plans are for the next two years, so I can decide whether it’s worth rooting for it.
That’s a fair perspective, but a $200 budget is actually more than enough for a highly productive workflow in Cursor. The key is its complete code synchronization, which provides so much context that you rarely need the most expensive models to get great results.
By using clearer prompts (prompt engineering) or outlining tasks in .md files, you can be even more efficient with top LLMs like GPT and Claude. Plus, with recent community rumors about a potential performance drop (“nerf”) in competitors like Claude Code, Cursor is in a very strong position. Many are highly productive even on the $60 Pro Plus plan, so a $200 budget definitely allows for unrestricted work.
thought Auto consumes Credits too? Last i checked it took like between 4 and 20 Credits (though supposedly it was bugged when it was introduced) back then, guess they switched to o3 max or something in Auto mode ?
But I honestly don’t understand how people make entire projects using such simple models. I have tasks where GPT-5-high runs for 20-40 minutes. And then I redo or continue with GPT-5-high or Gemini 2.5 Pro.
I hope that their projects are simple or boilerplate enough, and I’m not clumsy prompter.