Claude Code is expensive, reckless, and weirdly fun

When Anthropic announced their new CLI, Claude Code, I wasn't amused. In fact, I didn't see the point. Why use it in the CLI when we have it in a proper IDE like Cursor?

However, after reading some interesting reviews, I decided to give it a shot. Here's what I learned over the last week (and after spending a few dollars along the way).

Comparison with Cursor

Let's start by addressing the elephant in the room. Why bother using Claude Code at around $5 per session when you can pay $20 for Cursor and use it indefinitely^[1], while enjoying a fully-fledged IDE rather than a dull CLI?

In my experience, there are two distinct use-cases where each tool clearly shines.

Managing Big Contexts

When precise context is crucial, Cursor is the clear winner. Its ability to control exactly what code is fed into the model helps tailor responses precisely. Suppose your code follows a specific pattern—maybe you organize your API layer into routers, services, and schemas. If you want to create a new service, it’s helpful to open related files, add them to the context, and prompt the model with instructions like "follow the pattern defined in X, Y, and Z."

In Claude Code, however, there's no practical way to tag files, only to describe them verbally. This approach is suboptimal, as it either requires a very small codebase or perfect memory of file paths—something I rarely have or think I should have.

Vibe Coding

Ah yes, the term of the hour: vibe coding. It's essentially when you interact more than you read or write code. Development in this mode is somewhat reckless and, honestly, fun.

Claude Code surprisingly excels here. Operating outside the IDE helps get into the mood—you skim rather than deeply analyze code. While the loader runs, you can do something else. It makes mistakes but often self-corrects. You’re truly supervising, making coding feel as close to natural language as possible in 2025.

Pair it with a tool like Superwhisper, and the experience becomes even better. In vibe coding, precise prompts aren’t crucial—you simply open your mind and let the AI sort your thoughts.

It's genuinely fun: pressing a button to talk, standing up, walking around, mumbling ideas on how something should work or pitfalls to avoid, and shaping thoughts in your head while AI does the heavy lifting.

Of course, this approach can get expensive. Vibe mode tends to be imprecise—sometimes taking longer and inflating context, potentially confusing the model and leading to inefficiency.

Yet, there's a fascinating trade-off: this relaxed method is far less cognitively intense, which makes speed irrelevant. Like not trying to break personal records every gym session, there's a place in coding for simply focusing on consistent progress—and if it's enjoyable, even better.

When Does It Shine?

Claude Code excels at tackling boring tasks. For instance, I successfully set up Pytest in minutes—a chore I'd procrastinated on due to sheer lack of enthusiasm. Quickly, I had tests ready, documentation drafted, and a new CI workflow established.

It's also great for extending features, especially when you allow for some creativity or don't have strong opinions about the solution. Need a reading mode for your news aggregator? Let the model handle it. Server-side search? Give it a spin.

However, due to lower supervision by design, it suits smaller or less critical projects best. If strict adherence to guidelines or best practices is required, Cursor’s superior context management typically provides better results.

Pricing

Let’s be clear—it’s expensive. Perhaps not so much in absolute terms, but definitely in relative terms.

If someone fixes your problem for $5, that seems reasonable. Over a month, that's akin to having a dedicated intern for $100—not bad, right?

Yet, relative to developers accustomed to around $20 monthly for comprehensive tools (OpenAI Plus, Claude AI, GitHub Copilot, Cursor), it can feel steep.

The enjoyment factor partly relies on reckless experimentation, suitable primarily for side or non-critical projects. But this sharply reduces ROI—are you willing to spend $20–$30 on a weekend side project likely generating $0 revenue, just for fun?^[2]

Honestly, without my company’s carte blanche for experimentation, I'd likely use Claude Code far less (though I'd be less amused at work).

Is It the Future?

Probably not—and I don't particularly care. Currently, it's a valuable tool for self-contained tasks permitting creative freedom or mechanical chores where the implementation specifics don't matter.

There are obvious UX issues likely soon resolved—like unnecessary permission prompts for harmless tasks—but ironically, its simplicity contributes significantly to its charm.

Of course, underlying model quality matters greatly—and Claude 3.7 Sonnet (and its reasoning variant) are arguably the current SOTA for programming, greatly enhancing the overall experience.

I'll continue reporting my experiences. Until then, enjoy your vibe coding (if your wallet allows).

1. It's not really indefinitely as Cursor may throttle you, connection drops, there are fast and slow requests, etc. ↩︎
2. Not that I never regretted spending way more than that for things that are way less fun, but YMMV. ↩︎