Cursor has recently been making waves on social media because of its killer product experience. It’s so good, in fact, that I’ve switched from VSCode to Cursor. When Cursor first came out, it was quite crappy. But over the past two years, it’s really made great progress. It’s one of those rare gems that’s managed to hold its own against the big guys.
Cursor received angel investment from OpenAI in September 2023 and announced a Series A financing of $60 million in August 2024.
Compete with Github head-by-head
The company behind Cursor, Anysphere, was founded in 2022, with Cursor being their first product, launched in 2023. By the time Cursor hit the scene, the Coding Copilot field had already been developing for quite some time and had several major players.
Officially launched officially on VSCode in March 2022, Github Copilot had already reached 1.3 million paid users by January 2024, making it the largest Coding Copilot product. It’s still growing at a rapid pace.
Codeium shifted its focus to Coding Copilot products in 2022. They announced a Series C funding round of $150 million in August 2024 and currently has 700,000 free users. They’ve also hit an eight-figure ARR in the enterprise market.
Sourcegraph, primarily involved in code search, launched a new product line “Cody” at the end of 2023. They have not yet disclosed user data. Sourcegraph had previously disclosed a Series D financing of $125 million in 2021.
Launched in 2019, Tabnine has now received a Series B financing of $25 million. According to disclosed data, it had reached 1 million users and earned a revenue of $5.8 million in 2022.
Apart from the notable commercial success of Github Copilot, other competitors, despite substantial funding, have been lagging in development. In this context, a group of young people stepped up, proposing to launch an AI Native IDE built on VSCode, with OpenAI’s model, and taking on Microsoft (Github). They bravely challenged a giant’s mainstream product using his own technology. Despite the seemingly improbable narrative, these young innovators managed to secure $8 million in funding.
Game Changers
The key breaking point from a technical perspective is simple: Cursor found a better way to commercialize almost identical technology. This is a a textbook example of both Product-Market-Fit and Product-Model-Fit. The better productization is primarily reflected in one core aspect: AI’s ability to modify multiple parts of a single file or multiple files at the same time. You can refer to the details on Cursor’s official blog.
The rapid development of LLMs has unlocked more scenarios. The product experience that Cursor wanted to create could not be achieved during the reign of GPT-4 or even GPT-4-Turbo because the model was too slow. However, Sonnet 3.5 and Llama 3 appeared in the first half of 2024. Although these two models did not surpass GPT-4 in terms of intelligence, they performed better and Llama 3 provided application side with a normally intelligent, fine-tunable model for the first time. These two points may not mean much to the entire industry, but they are the key factors to unlock the product experience envisioned by Cursor. This is Product-Model Fit.
Changes in OpenAI HumanEval Evaluation Scores Over Time
Cursor’s product experience has its roots in Github Copilot, and you could say that Github helped Cursor achieve Product-Market-Fit. So, from Day 1, Cursor had formed a viable business model. Developers are more than willing to pay out of their own pockets for efficiency, even for minor improvements.
Large companies are very slow to react. It’s hard to imagine that Cursor is stealing users right under Github’s nose, and Github seems to have no reaction so far. Cursor’s current product experience is not something that only Sonnet 3.5 can achieve; GPT-4o can also do it. But why didn’t Github do it? Some say it’s because VSCode and Github are two different departments, and divisional silos. If anyone with some influence within Github wanted to do what Cursor is doing, they could potentially nip Cursor in the bud. But, such things often don’t happen in Silicon Valley. In a sense, this is also Product-Market-Fit.
The aforementioned points, I believe, are necessary conditions for Cursor’s success, but these conditions are the same for many startups. So why haven’t others succeeded? The key is that Cursor’s accurate product assumption, and then kept refining and waiting for technology to mature to make this assumption a reality. The right question is often more important than the solution. I’m left wondering, how much of this is strategic foresight and how much is pure luck?
Concerns Amid the Boom
If users can easily switch from Github Copilot to Cursor because of experience improvement, does it mean they can easily switch away? This raises questions about moat problem of products like Copilot. The real moat is VSCode, not Github Copilot or the current Cursor. VSCode has evolved from a simple editor to a platform. The reason users can easily switch from Github Copilot to Cursor is that they both rely on VSCode, and the user habits, experience, and features/plugins are all indentical. If Cursor cannot transfer VSCode’s existing experience, then it cannot convince users to give up VSCode with its current improvements. It is challenging to replicate the various plugins and expandable capabilities offered by the VSCode ecosystem. So both Github Copilot and Cursor are tools in the VSCode ecosystem. For developers, they are just enhanced hammers and do not have network effects or other moats. Also, Cursor’s case proves that the so-called data flywheel of Copilot products does not exist. Github’s early start did not result in advantages of product experiences such as efficiency, accuracy. The data you can get, the LLMs can also get, it’s already a part of the model.
If you’re following this field, you probably know that another IDE, Zed (zed.dev), has recently started to gain attention. From an engineering perspective, the complexity of Zed’s project is at least 10× of Cursor’s, as Zed built its IDE from scratch, completely independent of VSCode, and also integrated AI Copilot capabilities. However, based on current user adoption, Cursor’s user base is at least 10× of Zed’s. Indeed, right direction is always more important than hard work.
What Next?
In June, I attended a talk by the founder of Cursor in San Francisco. Their core philosophy is to create awesome tools that enable human developers into superhuman ones. This concept is quite different than current startup narratives.. I remember at that time, the sub-forum was titled Developer Agent, and Cursor is the only one discussing Copilot, while the rest were focused on the concept of Agent.
Is Copilot the path to becoming an Agent? There is still no common consent on this one, and I personally don’t believe so. Although Cursor hasn’t developed an Agent product yet, I surely believe they are cooking something here. Not long ago, OpenAI released the SWE-bench_Verified test set, currently recognized as the most authoritative benchmark in this realm. The competition is heating up, and it’s already intense, with many players who are quite different from those in the Copilot field:
SWE-bench_Verified tests AI on real-world software problems
The foundation of the current Agent technology is still in its infancy. Even though we (Gru.ai) currently have the highest score, it’s merely around 45, far from meeting commercial application standards. Based on history experience with the OpenAI HumanEval benchmark, we estimate that when the average score on this leaderboard is above 70, this field will have a commercially viable technology base. It’s fair to say that the Agent field hasn’t yet hit either of the two PMFs, with no validated product form and no mature technology.
Yet, this field continues to draw considerable interest. There is an increasing number of startups joining the game, with YC alone graduating at least 3 related teams. On one hand, everyone is building up their technological arsenal, and on the other hand, they’re seeking commercial viability, starting with solving small problems in software engineering, such as writing documentation, testing, etc. Of course, many have also raised doubts about this field—could it be too ahead of its time?
Pessimists may be right, but it’s the optimists who keep pushing forward. If you’re interested in this field, feel free to contact us at connect@gru.ai.
