Cursor Changed How We Write Code — Now Every IDE Is Scrambling to Catch Up

The fastest-growing SaaS company of all time was built by four people who never worked at a tech company. Michael Truell, Sualeh Asif, Arvid Lunnemark, and Aman Sanger were students at MIT when they founded Anysphere in 2022. They rejected offers from Big Tech, spent nearly a year building mechanical engineering tools nobody wanted, and then pivoted to the product that would make them all billionaires before 30: Cursor, an AI-native code editor.

In March 2026, Cursor surpassed $2 billion in annualized revenue, roughly doubling from $1.2B ARR just three months earlier. The company's Series D in November 2025 valued it at $29.3 billion — up from $400 million at its Series A fourteen months prior. That is a 73x valuation increase in just over a year. The round minted four new billionaires — the four co-founders — and drew capital from Accel, Thrive Capital, Andreessen Horowitz, and Google.

These are staggering numbers. But the most interesting thing about Cursor is not its growth trajectory. It is the fact that its own CEO went on stage at Fortune Brainstorm AI in December 2025 and warned the world not to trust the code his product generates.

"If you close your eyes and you don't look at the code and you have AIs build things with shaky foundations as you add another floor, and another floor, and another floor, things start to kind of crumble," Truell said.

That tension — between a product growing faster than any SaaS tool in history and a founder telling you to keep your eyes open while you use it — is the story of AI coding in 2026.

The Market Cursor Built

To understand Cursor's position, you need to understand the decision the founders made before writing a single line of code. They chose not to build a plugin. Every other AI coding tool in 2022 was an extension bolted onto an existing IDE — a sidebar, an autocomplete layer, a chat widget. Cursor's bet was different: fork VS Code entirely and rebuild the development environment with AI at its core.

The distinction matters architecturally. A plugin is constrained by the host IDE's APIs. A fork controls the entire surface — the editor, the file system, the terminal, the context window. That control enabled features that plugins could not match: Composer mode for multi-file editing with project-wide awareness, Agent Mode running up to eight parallel agents on a single prompt using git worktrees, and Background Agents that work on separate branches and open pull requests for human review.

By early 2026, Cursor had crossed 1 million users, including roughly 360,000 paying customers and over 50,000 enterprise seats across Fortune 1000 companies. The newest feature, Cursor Automations, introduces always-on agents triggered by codebase changes, Slack messages, PagerDuty alerts, or scheduled timers — turning the editor into an event-driven engineering platform.

The revenue trajectory puts this in context. Cursor went from $100M to $1.2B ARR in 2025 — an 1,100% year-over-year increase — then doubled again to $2B in roughly three months. Revenue was doubling approximately every two months at peak growth. No SaaS company has ever scaled from $1M to $500M in ARR faster.

The Competitive Pileup

Cursor did not create the AI coding category. GitHub Copilot, powered by OpenAI, launched in 2021 and currently claims over 20 million users and 1.3 million paid subscribers. It powers 90% of Fortune 100 companies and captures 49% adoption among developers already using AI tools. At $10 per month, it is the cheapest premium option in the market.

But Copilot is a plugin. It lives inside VS Code, JetBrains, Eclipse, and Xcode. That architectural choice gives it enormous distribution — it goes where the developers already are — but limits what it can do. Microsoft has been aggressively adding agent capabilities: agent mode launched in July 2025, and Copilot now integrates third-party agents from Anthropic and OpenAI. But it is playing catch-up on the full-IDE experience that Cursor pioneered.

The Windsurf saga is the cautionary tale of the cycle. Originally called Codeium, the company rebranded to Windsurf in 2024 when it shifted from code completion to a full agentic IDE. It built Cascade, a system handling multi-file edits autonomously, and reached $82M ARR with 350+ enterprise customers. In May 2025, OpenAI agreed to acquire Windsurf for $3 billion — what would have been its largest acquisition. The deal fell through when the exclusivity period expired in July. Google then executed a $2.4 billion reverse-acquihire, poaching CEO Varun Mohan, co-founder Douglas Chen, and key research leaders. Days later, Cognition AI — maker of Devin, the viral "first AI software engineer" — acquired what remained: IP, product, trademark, brand, and the remaining team. Cognition was subsequently valued at $10.2 billion.

One company. Three acquirers. Two months. The speed at which Windsurf was dismembered illustrates both the strategic value and the fragility of AI coding startups.

Meanwhile, Anthropic's Claude Code has emerged as a serious alternative with a fundamentally different philosophy. Launched in February 2025 and reaching general availability by May, Claude Code is terminal-native — it executes commands directly on your local machine, searches and edits files, runs tests, and pushes to GitHub. It surpassed $1 billion in annualized revenue by November 2025. Where Cursor wraps AI inside an IDE, Claude Code treats the entire development environment as its workspace.

Replit rounds out the field with a different bet entirely: an end-to-end cloud IDE where Agent 3 works autonomously for 200 minutes with built-in browser testing and self-fixing capabilities. The bet paid off — Replit's revenue jumped from $10M to $100M in nine months after launching their Agent.

The market these companies are fighting over is projected to reach $23.97 billion by 2030 at a 26.6% CAGR. CB Insights estimates the top three players capture over 70% of market share, and seven companies have already crossed the $100M ARR threshold.

The Study That Broke the Narrative

Every AI coding company sells the same story: developers are faster with AI. The numbers vary — 20%, 30%, 55% — but the direction is always the same. Then METR published a study in July 2025 that upended the entire narrative.

The setup was rigorous. Sixteen experienced open-source developers, each with an average of five years of experience on their respective projects, completed 246 tasks in a randomized controlled trial. Half the tasks were done with AI tools (Cursor Pro with Claude 3.5/3.7 Sonnet — frontier models at the time). Half were done without. The tasks were real: bug fixes, feature additions, and refactoring on mature codebases.

The result: AI tools made developers 19% slower.

Not faster. Slower.

The perception gap was the more damning finding. Before the study, developers predicted AI would make them 24% faster. After completing the tasks, they estimated AI had made them 20% faster. The measured reality was a 19% slowdown. The gap between belief and performance was 39 percentage points.

The study identified low AI reliability as the primary factor. Developers accepted less than 44% of AI generations. The time spent prompting, reviewing, and correcting AI-generated code exceeded the time saved by not writing it manually.

This does not mean AI coding tools are useless. Vendor-sponsored studies from GitHub, Google, and Microsoft — all companies that sell AI coding tools — found 20% to 55% speed improvements on scoped tasks like writing functions, generating tests, and producing boilerplate. The key difference is the task scope: AI excels at well-defined, bounded problems. It struggles with the ambiguous, cross-cutting work that occupies most of a senior developer's day.

Research from Faros AI added another layer. Their analysis found that AI coding assistants increase developer output but not company productivity. Delivery metrics — lead time, defect rate, deployment frequency — often remain unchanged even when individual output rises. The bottleneck migrates downstream to code review, QA, security audits, and integration testing. Developers produce more code. The organization does not ship more product.

METR announced in February 2026 that they are redesigning the experiment, noting that AI tools have improved significantly since early 2025. The next round of results will matter enormously for the industry's credibility.

From Vibe Coding to Agentic Engineering

The vocabulary of AI-assisted development has shifted faster than the tools themselves. In February 2025, Andrej Karpathy — OpenAI co-founder and former Tesla AI lead — coined the term "vibe coding" in a post that went viral. He defined it as "fully giving in to the vibes, embrace exponentials, and forget that the code even exists."

The term captured something real. A generation of developers — and a much larger cohort of non-developers — began building software by describing what they wanted in plain English and letting AI figure out the implementation. Collins English Dictionary named it Word of the Year for 2025. Merriam-Webster added it in March 2025 as "slang & trending."

Exactly one year later, Karpathy himself declared vibe coding "passe". His new term: agentic engineering — "agentic because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight."

The shift reflects the evolution of the tools. In early 2025, AI coding meant autocomplete and chat-based code generation. By early 2026, it means orchestrating multiple autonomous agents working in parallel across different branches, triggered by events, capable of opening pull requests and running tests without human intervention. The developer's role is shifting from writer to reviewer, from implementer to architect.

The numbers support this framing. AI now writes 41% of all code globally. 84% of developers use AI tools in 2026. Satya Nadella revealed that roughly 30% of Microsoft's code is now AI-generated, with a stated goal of reaching 80%.

Andrew Ng has pushed back on the vibe coding framing, arguing it misleads people into assuming developers just "go with the vibes." The more accurate description of modern AI-assisted development is closer to Karpathy's updated term: developers as supervisors, AI as the workforce, and judgment as the critical skill.

The Junior Developer Crisis

The most consequential impact of AI coding tools is not on productivity. It is on the talent pipeline.

A Stanford University study found that employment among software developers aged 22-25 fell nearly 20% between 2022 and 2025. Entry-level tech hiring decreased 25% year-over-year in 2024. A 2025 LeadDev survey found that 54% of engineering leaders plan to hire fewer juniors, as AI copilots enable senior developers to handle work that previously required additional headcount. Forrester forecasts a 20% drop in computer science enrollments and a doubling of time to fill developer roles.

The logic is straightforward and brutal. If a senior developer with Cursor can do the work of 1.5 developers, the headcount that gets cut is the junior hire. The tasks that juniors traditionally handled — boilerplate code, simple bug fixes, test writing, documentation — are precisely the tasks that AI tools handle best. The apprenticeship model that turned juniors into seniors is being hollowed out.

This creates a compounding problem. If companies hire fewer juniors today, there will be fewer experienced seniors in five years. The industry is optimizing for short-term efficiency at the potential cost of long-term capability. The developers who built the codebases that AI was trained on got their skills through years of writing code by hand. If the next generation skips that step, the quality of human oversight — the very thing that Cursor's CEO says is essential — degrades.

The counterargument is that the role, not the profession, is shifting. Juniors who can orchestrate AI tools, review generated code, and think architecturally remain valuable. But that requires a different kind of training than most computer science programs currently provide, and the transition period will be painful for the cohort caught in between.

The Code Quality Problem Nobody Wants to Talk About

Speed is the metric every AI coding company optimizes for. Quality is the metric they avoid discussing.

The data is uncomfortable. 48% of AI-generated code contains security vulnerabilities. Only 29-46% of developers trust AI code outputs as of 2026. These are not hypothetical risks. Cursor itself has been hit by CVE-2025-54135 and CVE-2025-54136 — remote code execution vulnerabilities via malicious repositories (dubbed CurXecute and MCPoison). Enterprise telemetry transmits commit information to Cursor servers, and for company subscription users, this telemetry cannot be disabled. CISOs are actively blocking Cursor adoption, demanding DLP plans, tenant isolation, and vendor SOC 2 certifications before approving even a pilot.

Cursor's own product quality has drawn criticism. The 2.1 release in November 2025 corrupted chat histories and worktrees, prompting prominent developer Theo to strongly advise against updating. Users have reported persistent file-saving failures, performance degradation on large codebases, and AI agents that change unrelated files without permission or provide false information about modifications made.

A January 2026 marketing stunt claiming to "vibe-code" an entire web browser was debunked on Hacker News. The FastRender project that was supposed to showcase AI capabilities showed an 88% job failure rate. The Register's headline: "Cursor shows AI agents capable of shoddy code at scale."

Then there is the pricing controversy. In mid-2025, Cursor changed its Pro plan from 500 fast responses plus unlimited slow responses to $20 worth of usage billed at API rates. Users reported running out of requests after just a few prompts with Claude models, with actual bills reaching $44 per month versus the advertised $20. Cursor publicly apologized for "unclear pricing changes." For a company approaching $2B in revenue, the pricing episode revealed the tension between growth and trust that runs through the entire AI coding industry.

The Shaky Foundations Warning

The most revealing statement about the state of AI coding in 2026 did not come from a skeptic or a competitor. It came from the person who has benefited more than almost anyone: Cursor's own CEO.

Michael Truell's warning at Fortune Brainstorm AI was specific and deliberate. He was not saying AI coding tools are bad. He was saying that the way many people use them — accepting generated code without review, building feature upon feature on unverified foundations — creates compounding technical debt that eventually collapses.

The metaphor of adding floors to a building with shaky foundations is precise. Each floor looks fine in isolation. The structural failure only becomes apparent under load, at scale, or when something unexpected happens. In software, that translates to security vulnerabilities, performance degradation, and bugs that are nearly impossible to trace because the developer who "wrote" the code never actually understood it.

This is the paradox at the center of the AI coding boom. The tools are powerful enough to let developers build faster than ever. They are not yet reliable enough to let developers build without looking. And the economic incentives — ship faster, hire fewer people, hit revenue targets — push relentlessly toward closing your eyes.

What Comes Next

The AI coding market is consolidating rapidly. The top three players control over 70% of market share. Seven companies have crossed $100M ARR. The projected market size of $23.97 billion by 2030 means there is room for multiple winners, but the window for new entrants is closing.

The technical frontier is moving toward full autonomy. Cursor's Automations, Claude Code's terminal-native architecture, and Replit's 200-minute autonomous sessions all point in the same direction: AI that does not assist developers but replaces discrete chunks of the development workflow entirely. Sam Altman predicted in 2025 that AI agents would move from completing multi-hour tasks to multi-day tasks. That timeline is compressing.

The question is whether the industry can scale the tooling without scaling the problems. The METR study will be rerun with improved models. The junior developer pipeline will either adapt or atrophy. The security vulnerabilities in AI-generated code will either be solved through better tooling or exploited at scale. And the companies building these tools will either solve the trust problem — proving that AI-generated code is reliably safe — or they will build a generation of software on foundations that even their own CEOs call shaky.

Cursor is the fastest-growing SaaS company in history. It is also, by its own founder's admission, a tool that requires constant human vigilance to use well. Both of those things are true simultaneously. The companies and developers who internalize that contradiction — who use the speed without surrendering the judgment — will be the ones who build things that last. The ones who close their eyes will build things that crumble.

The floor count is going up. The foundation has not changed.