Cursor changed its pricing model in mid-2025. Some developers saw bills jump significantly. Reddit threads filled with complaints. Headlines declared an "exodus."
But is that what's actually happening?
The honest answer: we don't fully know. Some developers are switching tools. Some are returning to simpler options. But calling it an "exodus" might be overstating what's really just normal market dynamics—people switching when prices change or tools don't meet expectations.
Here's what we can say with reasonable confidence, what's uncertain, and what context you need to evaluate the situation yourself.
What Actually Happened With Pricing
The Cursor Pricing Change
In mid-2025, Cursor shifted from request-based limits to a credit system. The company acknowledged the transition wasn't handled well:
Some users reported significant cost increases. How widespread? Hard to say. Reddit complaints are visible but not statistically representative. Users who had no problems rarely post about it.
Missing Context
We don't have data on:
- What percentage of Cursor users saw higher bills
- How many users actually canceled vs. complained but stayed
- Whether heavy users (higher bills) represent typical usage patterns
The Broader Pattern
Other tools also adjusted pricing around the same time—Claude Code, Copilot, Windsurf, Replit. This wasn't coordinated conspiracy. It reflects a market reality: AI coding tools cost money to run, and early pricing wasn't sustainable.
The Information reported Replit's margins swung between -14% and +36% in 2025. TechCrunch sources described some AI coding assistants as "massively money-losing businesses."
Translation: the cheap pricing was subsidized to acquire users. That's now ending across the industry.
What This Doesn't Mean
Higher prices don't automatically mean the tools aren't worth it. They might be. They might not be. It depends on your usage patterns, alternatives, and what you're getting in return. The value calculation is individual.
The Productivity Question
The METR Study: Real but Limited
In July 2025, METR published a study finding experienced developers were 19% slower with AI tools on tasks in familiar codebases.
This study gets cited constantly. But context matters:
| What the Study Found | What It Means | Limitations |
|---|---|---|
| 19% slower with AI tools | Point estimate for experienced devs on familiar codebases | 95% CI: -43% to +12%. Could be anywhere in that range. |
| Developers thought they were 20% faster | Perception doesn't match reality for this sample | Sample size was 16 developers. That's small. |
| 44% acceptance rate for suggestions | More than half of AI output was rejected | Doesn't tell us about the value of accepted suggestions |
The Deeper Issue: AI as Decision-Maker
Here's something the productivity debate often misses: AI coding tools aren't just autocomplete. They're making architectural decisions, choosing patterns, structuring code. As Yuval Noah Harari argues, AI represents something new—an entity that makes decisions autonomously, not just a tool that amplifies human choices.
When AI generates code you didn't anticipate, you have to evaluate decisions you didn't make. That's cognitively different from writing code yourself. It might explain why experienced developers—who have strong opinions about how code should work—experience friction with AI suggestions.
What the Study Doesn't Tell Us
- Unfamiliar codebases: The study tested experienced devs on codebases they knew. AI might help more with unfamiliar code.
- Junior developers: The study focused on experienced developers. Results might differ for juniors.
- Different task types: Boilerplate, documentation, and exploration might show different patterns than bug fixes.
- Learning curve: Developers had varying AI experience. Mastery might change outcomes.
The "Tool Fatigue" Phenomenon
Beyond pricing and productivity, there's another factor: exhaustion from constant change.
One developer captured the sentiment: "My 'default tool' may change again next month. It's both exciting and exhausting."
This is real. The AI coding tool landscape changes rapidly. New features, new models, new workflows. For some developers, that's energizing. For others, it's draining.
What We Actually Know vs. Don't Know
We know:
- Some developers are switching from Cursor to VSCode + Copilot
- Stack Overflow shows positive sentiment for AI tools declined from ~70% to ~60%
- Pricing has increased across most AI coding tools
- The tool landscape changes frequently
We don't know:
- Whether switching represents a significant trend or normal churn
- Whether sentiment decline indicates fundamental problems or adjustment to realistic expectations
- How many developers benefit from AI tools vs. don't
- Whether the METR findings generalize beyond 16 developers
Why Some Developers Are Choosing Simpler Tools
For developers who are switching back, the reasons tend to cluster around a few themes:
Predictable Costs
GitHub Copilot: $10/month, flat rate. You know what you're paying. For developers managing budgets—especially paying out of pocket—predictability matters.
Familiarity
VSCode is familiar. No new hotkeys. No new mental models. Integration with existing workflows without relearning.
Stability
Microsoft-backed tools are likely to exist next year. Startup tools might not. For some, that matters more than cutting-edge features.
Skill Maintenance
Some developers worry about skill degradation. If AI writes all your code, are you still learning? This concern—valid or not—leads some to deliberately use less AI assistance.
What Smart Developers Are Actually Doing
Rather than prescribe a "right" answer, here's what we're seeing experienced developers consider:
Measuring actual productivity: Not trusting feelings. The METR study showed developers perceived 20% speedup while experiencing 19% slowdown. Track completion times if you want real data.
Matching tools to tasks: AI might help with boilerplate but hurt with complex logic. Some developers use AI selectively rather than universally.
Budgeting for realistic costs: The $20/month era is ending. Budget for $50-100/month for serious AI coding assistance, or stick with lower-tier options.
Evaluating switching costs: Changing tools has costs too. Learning curves, workflow disruption, integration work. Sometimes staying put is the right choice even if another tool is theoretically better.
The Uncomfortable Questions
Here's what we don't know how to answer:
Is the productivity decline permanent or a learning curve? Maybe developers get faster with AI over time. Maybe they don't. We don't have longitudinal data.
Are the frustrated developers representative? People who complain are visible. People who quietly use AI tools successfully are invisible. We might be hearing from a vocal minority.
Is "exodus" the right frame? People switch tools constantly. They switched from editors to IDEs, from IDEs to VSCode, now from VSCode to AI-enhanced tools and sometimes back. Is this "exodus" or normal market behavior?
The Bigger Picture
AI coding tools represent something genuinely new: systems that don't just amplify our coding abilities but make coding decisions themselves. The friction developers experience might not be a bug to be fixed but a feature of working with autonomous decision-makers.
When you review AI-generated code, you're evaluating choices you didn't make. That's different from reviewing code you wrote yourself. Whether that's worth the tradeoff depends on individual circumstances—and we're all still figuring this out.
What This Means Going Forward
Prices will likely stabilize higher. The subsidized era is ending. Expect $50-100/month to become normal for premium AI coding tools.
Consolidation is coming. Tools with unsustainable economics will raise prices dramatically, get acquired, or shut down. Tools backed by major companies (Microsoft, JetBrains, Google) have advantages in sustainability.
Expectations are adjusting. The "10x productivity" hype is fading. More realistic assessments—AI helps with some tasks, hurts with others, depends heavily on context—are emerging.
The debate continues. We're still in early days. The developers who figure out when AI helps and when it doesn't—rather than using it universally or rejecting it entirely—will probably come out ahead.
The Bottom Line
Some developers are switching from AI-enhanced IDEs to simpler tools. Prices have risen. The METR study suggests AI might slow experienced developers down—but the sample was small and the confidence interval was wide.
Is this an "exodus"? Maybe. Or maybe it's normal market dynamics as pricing adjusts and expectations reset.
The honest answer: we're all still figuring this out. If your current setup works for you, that's data. If it doesn't, switching is reasonable. But be skeptical of anyone—including us—claiming to know definitively whether AI coding tools are net positive or negative.
The question isn't "are AI tools good or bad?" It's "under what conditions do they help, and under what conditions do they hurt?" We don't have complete answers yet. Anyone who claims otherwise is selling something.