On January 5, employees at Cursor returned from the holiday weekend to an all-hands meeting with a slide deck titled “War Time.”

During the break, employees playing with Anthropic’s latest model, Opus 4.5, had experienced an uncomfortable realization: its coding abilities had advanced to the point where developers no longer needed to review every line of output. Instead of collaborating with an AI assistant inside Cursor’s code editor, developers could issue high-level instructions to autonomous agents and receive back completed features — sometimes, even the finished product. And that was a problem.

Cursor was built on a different premise. CEO Michael Truell described it to Forbes in 2024 as a kind of “Google Docs for programmers,” a collaborative editor where humans and AI refined code together.

But if the AI doesn’t need a human collaborator, why bother with the editor? If writing and editing code line by line was no longer central to a programmer’s workflow, Cursor’s central product thesis was suddenly in question.

At the all-hands, Cursor leadership warned that the months ahead would be turbulent ones. Projects might be scrapped, priorities shifted. The company’s new mandate was labeled “P0 #1”—priority zero: “Build the best coding model.”

Not the best wrapper. The best model. Call it a vibe shift. Inside Cursor, it felt like a reckoning.

Which is what makes this moment so jarring. Until recently, Cursor seemed nearly unstoppable. The company began 2025 with roughly $100 million in annualized revenue. By November, that figure had surpassed $1 billion. Its latest financing round valued the company at nearly $30 billion, minting its four cofounders as billionaires and placing Cursor among the top 20 most valuable private companies in the world.

But in the fast-moving world of AI, perceived momentum can appear – or evaporate – overnight.

By February, after Anthropic released an even more advanced version of Opus, X started to fill up with startup founders claiming their teams had ditched Cursor, arguing that model makers like Anthropic and OpenAI would absorb the coding layer themselves.

“Most of the companies I mentioned…their view is that Cursor is obsolete today,” Insight Partners founder Jerry Murdoch said on the 20VC podcast last week.

But the numbers throw that narrative into question. According to a source familiar with Cursor’s finances, annualized revenue has now crossed $2 billion, doubling within three months. And data from corporate credit card companies Ramp and Brex show continued revenue growth through February, though Ramp says Cursor’s adoption rates among businesses buying AI products are declining slightly. It’s not clear yet whether Claude Code’s huge boost will ultimately impact Cursor’s growth.

Inside the company, Cursor’s leadership knows that the future of software development does not involve writing lines of code. To respond, they’ve been building the research horsepower to outmaneuver Anthropic and OpenAI in releasing the best coding models by publishing research and harnessing vast troves of proprietary data. And they started prioritizing contracts with large enterprises, which can be more stable than consumer subscriptions.

For now, Cursor’s continued growth comes with a big dose of anxiety. Inside the startup, revenue tracking became so distracting that the company stopped reporting daily figures in its #numbers Slack channel, according to people familiar with the decision.

Founded in 2022 by four MIT friends, Cursor started out by building models to help mechanical engineers to design physical parts. But the cofounders had no expertise in the area. They pivoted and landed on their breakout product: a code editor that exploded in popularity. The company’s super fast coding models eventually helped kick off the “vibe coding” phenomenon — where people could write (or speak) plain English prompts to AI models to code entire web apps.

Cursor’s founders, and a large portion of its 400 employees, are in their mid twenties, and the startup feels less like a company and more like an elite college campus. Employees slip off their shoes before entering the office. Staff regularly work past midnight, shower onsite and live within a few blocks’ radius of the office.

A year ago, Cursor was known as the startup whose product was so viral, it had crossed $100 million in annualized revenue with only 20 employees and no sales staff. Its rapid growth attracted attention from VC heavyweights like Accel, Andreessen Horowitz and Thrive Capital. And it has gotten front-of-the-line access to some of the best models. In 2025, Anthropic gave Cursor early access to its models and used its feedback to improve the models’ capabilities in a sort of “co-opetition.”

As the company has grown to roughly 400 employees, it has taken over four buildings in the North Beach neighborhood, and even commandeered the ad space in the bus stop between its offices for a poster listing employees’ names: Cursor’s version of scrawling names on a bathroom wall.

Early last year, Anthropic called Cursor, then its largest customer, with a preview of a new product called Claude Code, a command line tool with a minimal interface that allowed developers to quickly deploy armies of coding agents.

At first glance, it didn’t look like it would compete directly with Cursor’s code editor. But that’s no longer the case. Claude Code crossed $1 billion in annualized revenue within six months and hit $2.5 billion last month, surpassing Cursor. Meanwhile, OpenAI has been pushing forward in the same direction. After relaunching its coding agent Codex in April 2025, CEO Sam Altman said its app was downloaded over 1 million times in its first week.

Startup founders told Forbes the shift is profound. Instead of writing code line by line, many developers now orchestrate agents—assigning tasks, reviewing outputs, coordinating multiple parallel processes.

“It is the biggest, most fundamental change in software development since the beginning,” says Andrew Hsu, cofounder and CTO of AI language tutor app Speak. The company’s team of 50 engineers are all using coding agents (mostly Claude Code, but also Codex in some instances) to ship features in a matter of weeks instead of months. Cursor still plays a role in reviewing changes, he said, but it’s a diminishing one.

Cursor’s internal values include a blunt directive: “Delete the product,” an acknowledgement that the company’s future lies in coding agents akin to Claude Code and Codex. Last week, Cursor announced major updates to its own “Cloud Agents” product. Now, multiple agents can work on different tasks simultaneously within their own dedicated workspace and record their work.

Internally, Cursor’s leadership is betting that enterprises will value a product that’s not tied to just one model provider, an increasing concern for developers as models capabilities improve by the day and the scale can tip in anyone’s favor.

At the same time, Cursor is also trying to reduce its dependence on Anthropic and OpenAI. Its thesis is that a smaller, specialized coding model—trained on its proprietary data—can compete effectively even as its competitors invest in ever-larger frontier models. Roughly 20 AI researchers now work on the company’s Composer models, according to sources familiar with the company. They are built on powerful Chinese open source models like DeepSeek, Kimi and Qwen, then modified through additional training, and reinforcement learning using Cursor’s own data. The efforts have paid dividends: Composer 1.5 is fast, the second-most popular model on the platform and is significantly cheaper for Cursor to run than paying for Anthropic’s large models. It’s still pricey for developers to use: Composer 1.5 costs $3.5 per million tokens on input whereas OpenAI’s GPT-5.3 Codex costs $1.75 within Cursor, according to its website.

Cost remains an ever present challenge. Cursor’s larger rivals are willing to subsidize aggressively. According to a person familiar with the company’s internal analysis, Cursor estimated last year that a $200-per-month Claude Code subscription could use up to $2,000 in compute, suggesting significant subsidization by Anthropic. Today, that subsidization appears to be even more aggressive, with that $200 plan able to consume about $5,000 in compute, according to a different person who has seen analyses on the company’s compute spend patterns.

Cursor also subsidizes some users, though it appears it doesn’t do so as much as Anthropic. Cursor has negative margins for consumer subscriptions, but its business plans operate on positive margins, according to a person familiar with its finances. Businesses that use Cursor can use the Teams plan, which is targeted at startups and is easy to cancel, or negotiate an enterprise contract, which is targeted at larger organizations.

Expanding its enterprise business is one path toward stability. Enterprise contracts are slower to close but less prone to churn. Cursor has only ever lost one or two enterprise customers, a person familiar with the business said. But those coveted enterprise contracts have historically made up a small portion of Cursor’s business: as of November last year, only 13.6% of Cursor’s annualized revenue came from enterprise contracts, according to documents seen by Forbes. Today, about 60% of Cursor’s revenue comes from businesses, according to a person familiar with Cursor, though Forbes couldn’t determine what portion of that is enterprise plans.

The company’s staff now reflects the focus on enterprise: half are focused on go-to-market functions. The sales team has closed contracts with large customers, including Meta and Nvidia, according to people familiar with its plans.

Pressure is building. In February, more than 90 employees at mortgage servicing startup Valon canceled their Cursor subscriptions. The reasoning was blunt: they didn’t need the editor anymore. Instead, they shifted to Claude Code’s powerful agents to fully automate work end-to-end—migrating data between systems, fixing bugs—tasks CEO Andrew Wang says were getting done “10x faster.”

While the software development world is grappling with the impact of autonomous coding, an even newer paradigm around multi-agent systems is emerging: imagine a single developer operating a team of dozens of agents, each of which have their own role and function almost like human teammates might.

Now, Cursor is trying to figure out the best way to build a tool that could manage hundreds of agents working together simultaneously, something they internally call “grind mode.” There are complex issues to work out. They need to figure out the best way to assign specialized roles to each agent. Sometimes the agents, when they see they have so many coworkers, become lazy and underperform — just like humans.

© Forbes