In January of this year, the "AI programming star" Cursor, known for its "limitless potential," discovered that the market might be gone.

The most dazzling unicorn in the AI programming tools sector, Cursor, is facing a strategic transformation that could determine its survival.

On January 5th, Cursor employees returned from holiday to a company-wide meeting themed “War Time Status.” The trigger for this urgent gathering was a troubling technological reality: AI models’ programming capabilities have advanced to the point where human line-by-line code review is no longer necessary. Developers can directly issue high-level commands to autonomous agents and receive complete functionalities, fundamentally shaking the core product logic of Cursor’s “human-AI collaborative code editor.”

The company immediately set its highest priority task as “building the best programming model”—not just the best encapsulation tool, but the model itself. This shift occurred during Cursor’s seemingly unstoppable peak. The company’s annualized revenue was about $100 million in early 2025, surpassing $1 billion by November, with a recent funding round valuing it close to $30 billion.

A sudden change in market dynamics is forcing Cursor to rewrite its business logic: shifting from consumer subscriptions to enterprise contracts, moving from reliance on external models to developing proprietary programming models, and transforming from a code editor to a multi-agent collaboration platform. Whether this transformation succeeds will determine if this three-year-old company can maintain its leading position in the AI programming race.

A “War Mobilization” Triggered by the Holidays

The January 5th meeting was prompted by employees’ firsthand experience with Anthropic’s latest model, Opus 4.5, during the holidays.

This model’s programming ability has evolved to a new level: developers no longer need to collaborate with AI assistants line-by-line within the editor. Instead, they can issue high-level commands to autonomous agents and receive complete modules, sometimes even finished products.

This poses a direct threat to Cursor. CEO Michael Truell previously described Cursor in 2024 as “the Google Docs for programmers”—a collaborative editor where humans and AI refine code together. But if AI no longer needs human collaborators, the value of the editor itself is at risk.

At the all-hands meeting, Cursor’s management warned that the coming months will be turbulent, projects may be halted, and priorities will be reshuffled. The company’s top priority was labeled “P0 #1,” with a single goal: to build the best programming model. Internally, this moment was called a “liquidation.”

From Zero to a Hundred Billion: The Rapid Rise

Founded in 2022 by four MIT alumni, Cursor initially focused on helping mechanical engineers design parts. However, due to a lack of domain expertise, it quickly pivoted and eventually launched a market-breaking code editor product.

The company’s rise has been remarkable. Just a year ago, with only 20 employees and no sales team, Cursor reached $100 million in annualized revenue. Its viral product attracted top-tier VCs like Accel, Andreessen Horowitz, and Thrive Capital. By 2025, Anthropic listed Cursor as its largest customer and granted it priority access to its models, creating a “co-opetition” relationship.

By November 2025, Cursor’s annualized revenue exceeded $1 billion, with a recent funding valuation approaching $30 billion, ranking it among the top 20 most valuable private companies globally. The company expanded to about 400 employees, occupying four buildings in San Francisco’s North Beach district, even turning bus stop ads between offices into employee roster displays.

Despite doubts about its technological trajectory, Cursor’s financial data continues to grow explosively. According to Forbes citing insiders, Cursor’s annualized revenue doubled in three months and has now surpassed $2 billion. Data from corporate credit card companies Ramp and Brex also show continued revenue growth in February, although Ramp noted a slight decline in enterprise AI product adoption.

However, competitors are expanding just as rapidly. Anthropic’s Claude Code surpassed $1 billion in annualized revenue within six months, reaching $2.5 billion last month, overtaking Cursor. OpenAI CEO Sam Altman stated that after the re-launch of its coding agent Codex in April 2025, the first-week downloads exceeded 1 million. In this fast-changing AI landscape, market momentum can reverse at any time.

“Dimensionality Reduction” Attacks from Competitors

The crisis at Cursor stems from a fundamental shift in AI programming paradigms.

Anthropic had already announced plans for Claude Code early last year— a command-line tool with a minimalist interface that allows developers to quickly deploy numerous programming agents.

Initially, this product did not seem to directly compete with Cursor’s code editor. But the situation has changed dramatically. Several startup founders say this shift is profound.

Andrew Hsu, co-founder and CTO of AI language learning app Speak, said: “This is the most significant and fundamental change in software development history.” His company’s 50 engineers have all shifted to using programming agents (mainly Claude Code, with some scenarios using Codex). Features that once took months to develop can now be delivered in weeks, and Cursor’s role in their workflow is diminishing.

In February, mortgage startup Valon, with over 90 employees, canceled its Cursor subscription and switched to using Claude Code’s agents for end-to-end automation, including data migration and bug fixing. CEO Andrew Wang said task completion speed increased by “10 times.”

Jerry Murdock, co-founder of Insight Partners, stated plainly on the 20VC podcast: “Most of the companies I mentioned… their view is that Cursor is already outdated today.”

Two Paths to Breakthrough: Proprietary Models and Enterprise Market

Faced with pressure, Cursor is seeking breakthroughs along two paths.

In developing proprietary programming models, Cursor has about 20 AI researchers working on its own Composer series models. These are based on open-source Chinese models like DeepSeek, Kimi, and Qwen, further trained and fine-tuned with Cursor’s proprietary data through reinforcement learning.

The core idea is: a small, specially trained model can effectively compete with large, cutting-edge models in terms of cost and performance. Currently, Composer 1.5 has become the second most popular model on the platform, with significantly lower operating costs than Anthropic’s large models.

However, cost pressures remain intense. According to media reports, an insider estimated that internal calculations show Anthropic’s subsidies for Claude Code are aggressive— a $200/month subscription can consume about $5,000 worth of compute. Cursor also subsidizes some users, with consumer subscriptions operating at a negative gross margin, while enterprise plans are already profitable.

On the enterprise front, Cursor is accelerating its focus on large clients. Although enterprise contracts take longer to sign, customer churn is very low—an insider said Cursor has only lost one or two enterprise clients so far. Currently, about 60% of revenue comes from enterprise customers, with sales teams already signing deals with giants like Meta and Nvidia. The internal structure has also shifted, with half of the staff now dedicated to market expansion.

Compared to November last year, when enterprise contracts accounted for only 13.6% of annualized revenue, this rapid increase reflects a clear strategic shift.

“Deleting Products”: Betting on Multi-Agent Future

A straightforward internal directive at Cursor is “delete products”—a sober acknowledgment of its future: code editors will eventually give way to programming agents.

Last week, Cursor announced major updates to its “Cloud Agents” product, enabling multiple agents to work simultaneously in separate workspaces and record their processes. The company is exploring a “grind mode”—a setup where a single developer can dispatch hundreds of agents to collaborate.

This vision faces complex technical challenges: how to assign dedicated roles to each agent, and how to prevent efficiency drops when agents perceive “many peers”—a phenomenon similar to free-riding in human teams.

Cursor’s leadership believes that a key differentiator is: enterprise clients will increasingly value products that do not rely on a single model provider, especially as model capabilities fluctuate and the landscape shifts rapidly. Whether this judgment proves correct will largely determine if Cursor can hold its ground in the AI programming war.