Article | LetterAI
$26 billion is the latest valuation given to AI programming company Cognition by the capital markets.
Last September, Cognition AI had just crossed the $10 billion valuation threshold, making it already resemble a Silicon Valley legend.
Chinese, Olympiad, Harvard, MIT, dropped out to start a company, AI Agent... Each label is compelling enough; Cognition is undoubtedly one of the most story-driven companies in the AI programming space.
Now, this story has been pushed forward significantly by the capital markets.
According to Bloomberg, Cognition AI, the company behind Devin, has raised over $1 billion in new funding, bringing its post-money valuation to $26 billion. The round was co-led by Lux Capital, General Catalyst, and 8VC, with participation from Ribbit Capital, Atreides Management, Founders Fund, and others. Cognition has officially confirmed this funding round and the new valuation.
In other words, just over eight months since its last valuation of $10.2 billion, Cognition's valuation has increased to 2.5 times its previous value.
01 What capital buys isn't just an AI programmer
The lead investor in this funding round is highly representative.
Lux Capital is a highly recognizable hard tech fund in Silicon Valley, long focused on investing in cutting-edge science, deep tech, AI, robotics, aerospace, defense, and computing infrastructure—projects that are considered “highly technical.” On its own investment page, Cognition is categorized under the direction of “increasing productivity efficiency + infrastructure + computer science.”
It could be said that Lux Capital invested in Cognition because it sees the potential for Cognition to make AI agents into infrastructure for software engineering.
General Catalyst is focused on opportunities for AI to transform enterprise processes. The firm is not just a traditional venture capital company; on its website, it describes itself as a “global investment and transformation company,” and in recent years, it has strongly emphasized transformation—using capital, operations, and corporate relationships to drive change in traditional industries and large institutions.
In addition to Cognition, General Catalyst has also significantly increased its investment in Anthropic, participating in multiple large funding rounds over the past year.
8VC, also a lead investor, carries the imagination of government and large enterprise adoption. The firm has long bet on enterprise software infrastructure within complex organizations, and Cognition’s client list already includes government or public-sector entities such as the U.S. Army, U.S. Navy, and NASA. 8VC’s participation in the lead investment validates Cognition’s narrative.
In addition to the three lead investors, Cognition’s existing shareholder Founders Fund is also increasing its investment. The $400 million round in 2025 was led by Founders Fund, resulting in a post-money valuation of approximately $10.2 billion. Founded with participation from Peter Thiel, the firm is known for its aggressive investment style, favoring technology companies with the potential to reshape entire industries—such as SpaceX, Palantir, Anduril, Stripe, and OpenAI.
Lux Capital has long bet on hard tech and advanced computing, General Catalyst excels in enterprise software and large-scale institutional transformation, 8VC brings expertise in enterprise software and government markets, and Founders Fund is one of Cognition’s existing investors. The presence of these types of capital on Cognition’s investor list demonstrates that investors no longer view Cognition merely as a developer tool company, but as a potential candidate for the next-generation software engineering infrastructure.
A post-investment valuation of $26 billion fully demonstrates market confidence, and the primary reason capital is willing to continue driving up the price is growth.
Cognition has presented strong commercialization data: enterprise usage has grown more than tenfold this year, and revenue run-rate has increased from $37 million in May last year to $492 million today, with enterprise adoption of Devin maintaining 50% month-over-month growth over the past six months.
Although $492 million is not confirmed annual revenue but rather an annualized run-rate based on current revenue trends, this trajectory is still alarming. Investors can already see that enterprise customers are genuinely paying and using the service, with usage rapidly increasing—a remarkable achievement for a company founded in 2023.
The AI programming track is indeed thriving. Code, issues, tests, PRs, and documentation are inherently digital work objects; whether tasks are completed can be verified through testing, code reviews, and deployment outcomes.
For enterprises, software teams always have endless tasks, each of which is time-consuming and expensive (at least in terms of senior engineers’ hourly rates). If AI agents can reliably take over a portion of well-defined, repetitive, and verifiable software engineering tasks, they become a form of engineering capacity that companies are willing to pay for.
Behind the $26 billion, capital is truly betting on one insight: software development is becoming the earliest large-scale enterprise procurement scenario for AI agents.
02 After Devin went viral, reality brought a cold dose of truth
Cognition first gained widespread attention through an idea that seemed extremely bold at the time.
Before Devin, AI programming tools were mostly limited to the role of "assistants." GitHub Copilot helps programmers complete code, ChatGPT and Claude can explain errors and generate functions, and Cursor integrates AI directly into the editor, allowing developers to write and modify code in real time.
But Devin took a major step forward. It is directly defined by Cognition as an "AI software engineer"—users simply describe their needs in natural language, such as developing a website, building an application feature, or fixing issues in a codebase, and Devin will autonomously break down the task, write code, fix bugs, and continue until the project runs successfully.
In March 2024, when Cognition released its Devin demo, the entire developer community was electrified. It was promoted as the world’s first AI programmer and became one of the landmark products that truly brought the vibe coding movement into the mainstream.
The company’s founders also bring their own story: three Chinese founders—Scott Wu, Steve Hao, and Walden Yan—all came from the competitive programming community and collectively won five IOI gold medals. They are not traditional business-type founders; rather, they are more like a group of the most skilled young coders trying to train something else that can code.
After Devin's launch, the company quickly gained support from top-tier venture firms such as Founders Fund, Khosla Ventures, and Bain Capital Ventures, and its investor lineup rapidly took shape. Enterprise customers also began to emerge, with names like Goldman Sachs, Citigroup, and Ramp increasingly associated with Devin.
In July 2025, when Goldman Sachs introduced Devin, Fast Company’s headline literally read, “Goldman Sachs’ New AI Software Engineer Never Sleeps.” This highlights exactly what makes agents most compelling to enterprises: they can operate 24/7, require no shifts, and never pause for nights, weekends, or time zones.
That was Cognition’s earliest moment of glory: a young team, a group of Chinese founders with backgrounds in informatics competitions, an AI agent claiming to complete software development tasks end-to-end, backed by top-tier venture capital and major clients. Put together, it was almost the standard opening of a Silicon Valley AI legend.
When a tree stands out in the forest, the wind will surely break it; when a story is told too perfectly, problems will follow.
In the early days, Devin’s breakout was largely built on the company’s demo. As external developers began to analyze it frame by frame and test it in real use, skepticism emerged. Some believed the demo had been carefully edited to omit processes that made it appear less flawless—for instance, certain segments were questioned as being staged, where Devin allegedly created a bug and then fixed it itself, yet the presentation made it seem as though the task was completed smoothly all along.
Devin thus became embroiled in a period of controversy over "misrepresentation"—its promotional tone suggested that AI was already capable of operating independently, but real-world engineering environments are far more complex than demos.
Software development has never been just about writing code—it involves understanding requirements, making architectural judgments, retaining context, adhering to team conventions, and countless implicit constraints that never make it into issues. An agent running successfully doesn’t mean it’s always moving in the right direction; generating code doesn’t mean that code is worthy of merging.
After Devin's official launch, the disparity became even more apparent.
It had a high initial price, starting at $500 per month. But its performance didn’t seem to justify such a high cost: Answer.AI tested Devin for a month, assigning it 20 real engineering tasks, resulting in only 3 successes, 14 failures, and 3 uncertain outcomes.
The biggest issue is not just the high failure rate, but also the unpredictability of failures.
Some tasks aren’t particularly complex, but Devin still gets stuck in dead ends; some tasks are inherently unfeasible, yet it continues attempting them; sometimes it generates overly complicated, hard-to-maintain code, ultimately forcing engineers to spend more time reviewing and cleaning up.
And on top of that, it’s so expensive.
Cognition also recognized that the $500-per-month threshold was too high. In April 2025, Cognition launched Devin 2.0, reducing the starting price from $500 per month to $20 and introducing a more flexible pay-as-you-go pricing model.
But lowering prices isn't a cure-all; a tool designed to improve efficiency shouldn't end up causing people to waste even more time and energy—it's hard to justify.
The core tension in the early stages of autonomous agents lies here: the more the AI resembles an independent engineer, the more users need to trust it—but the more it runs autonomously as a black box, the bigger the trouble if it goes off track.
Devin promised to “hand over the task to me,” but many real engineering tasks aren’t suitable to be fully delegated this early. An Agent running independently for a long time and then handing you a PR sounds advanced; however, if the PR quality is inconsistent, the engineer’s review cost actually increases.
Interestingly, in this contrast, it was Cursor that reaped the first wave of genuine developer benefits.
Since Cursor never claimed to replace programmers, its approach is more gentle and aligns better with real workflows: AI assists alongside by modifying code, explaining errors, refactoring files, and generating tests, while developers remain in the editor—like a driving school car with dual brakes, allowing you to stop whenever something feels off.
If Cognition's story ended here, it might have become just another "traffic" company lifted by the AI hype and brought back down to earth by real-world user experiences—and none of what followed would have happened.
But as mentioned earlier, reality is often more complex, and the AI programming赛道 has not stood still.
After Devin sparked imagination around "AI software engineers" and Cursor demonstrated that developers still crave a sense of control, major foundational model companies like OpenAI, Google, and Anthropic have begun accelerating efforts to integrate coding capabilities into their own products and platforms.
On one side, a more controllable IDE pathway is rapidly expanding; on the other, model giants are moving down to consume the application layer. For Cognition to survive, it must make some changes.
At that very moment, it "found" the treasure left behind by Windsurf.
03 Grasp both hands, and strengthen both
The Windsurf rivalry can be considered one of the most dramatic events in the AI coding tools landscape of 2025.
At the time, Windsurf was already a highly prominent company in the AI IDE space. It had initially attracted interest from OpenAI, and the two parties engaged in prolonged discussions regarding an acquisition, leading many outsiders to believe the deal was all but finalized.
However, this transaction ultimately did not materialize, partly due to the complex partnership between OpenAI and Microsoft. At the time, Microsoft held broad licensing rights over OpenAI’s technologies and products, and Microsoft itself owned GitHub Copilot, a key competitor in the AI programming space. Windsurf feared that if acquired by OpenAI, its own technologies and products might become entangled in OpenAI’s licensing framework with Microsoft, indirectly flowing to potential competitors.
Just as OpenAI retreated, Google swiftly stepped in.
Google secured a non-exclusive license to Windsurf technology for $2.4 billion, along with Windsurf’s CEO Varun Mohan, co-founder Douglas Chen, and several key R&D personnel, who joined Google DeepMind.
It was a Friday, and things happened quickly: Google took away the founders and部分核心技术授权, OpenAI failed to complete the acquisition, and Windsurf’s original company entity, products, brand, customers, and 250 employees were left in a very awkward position.
At this moment, Cognition makes its entrance.
The incident occurred on Friday, and by Monday, it announced the acquisition of Windsurf's remaining assets, including the Windsurf IDE product itself, intellectual property, trademarks, brand, enterprise customer base, user data, and most of the remaining team members.
This step was nearly the key to Cognition's later return to the table, as it filled the very gap Devin lacked most: the developer entry point.
After acquiring Windsurf, Cognition significantly accelerated its commercialization. At the time of acquisition, Windsurf already had $82 million in annual recurring revenue (ARR) and over 350 enterprise customers; Cognition later disclosed that this acquisition more than doubled the company’s ARR, and within seven weeks post-acquisition, the combined enterprise ARR grew by over 30%.
Previously, Devin represented a more aggressive approach, aiming for users to hand tasks over to a cloud-based agent that would plan, execute, and debug itself before delivering the final result. However, the rise of Cursor has shown that developers are not necessarily willing to fully delegate tasks from the start. They prefer to stay within their editors, watching the AI modify code step by step, ready to take control and correct course at any moment.
The addition of Windsurf brings an IDE to Cognition, finally giving Cognition more than just the product form of "assigning tasks to AI."
It now walks on two legs: one is Devin, handling cloud-based asynchronous tasks suited for engineering work that can be broken down, verified, and delivered as PRs; the other is Windsurf, providing an IDE entry point that enables developers to work alongside AI directly in their codebase, covering everyday development scenarios similar to those of Cursor.
If users aren’t comfortable handing over full control to AI, let the AI return to the editor as a controllable assistant; if a company truly has a large volume of clear, repetitive, and verifiable engineering tasks to handle, let Devin act as a “formal employee” and take over some of the work in the background.
Cognition no longer solely pursues an all-powerful, autonomous AI programmer capable of independently completing all tasks; it now addresses two real-world needs in software engineering.
This contrasts sharply with the recent controversy surrounding Antigravity 2.0: Google initially started with an IDE, but after the Antigravity update, it directly shifted to an Agent Manager interface, moving from a controllable IDE collaboration model to a black-box agent scheduling system. The direction is ambitious, but it risks encountering the same issues Devin faced early on.
Individual developers often choose tools based on feel, efficiency, price, and experience—if a tool is hard to use, it’s quickly abandoned. But enterprises buy processes and capacity; if an Agent can integrate into their existing engineering system and consistently deliver results on certain tasks, it has a chance to become a budget item.
The most noticeable change in Cognition's narrative here is this.
Early Devin was like an AI programmer in the spotlight, proving it could write code just like a human programmer (and without needing rest). Later, Cognition became more like a seller of an enterprise engineering automation system: Devin handles asynchronous execution, Windsurf manages the development entry point, and enterprise customers integrate them into their own software development workflows.
According to a recent interview with TechCrunch on May 29, CEO Scott Wu is clearly pulling Devin back from the narrative of being a "replacement for programmers." When asked whether Devin can replace a mid-level programmer, his answer was, "Yes and no."
He emphasized that Cognition never intended to position Devin as a replacement for humans; the team itself consists of programmers who do not wish to see programmers lose their jobs. He noted that Devin’s capabilities vary depending on the task, generally falling between junior and mid-level engineer levels. It is better suited for handling long-tail maintenance tasks that many programmers dislike, such as upgrading legacy software or migrating platforms, thereby freeing engineers to focus on more creative work.
The combination of the two legs perfectly avoids the shortcomings of a single product. Relying solely on Devin might appear too aggressive, causing users to worry about the uncontrollability of autonomous agents; relying solely on Windsurf, on the other hand, would lead to direct competition with products like Cursor, Copilot, Claude Code, and Codex. But with Devin plus Windsurf, Cognition has a more complete story: it can serve both developers in their daily coding scenarios and enterprises in delegating tasks to agents.
In the latest funding round, the data presented by Cognition also shows that its story is being validated by the market.
The company said enterprise usage has grown more than tenfold this year, with a revenue run-rate reaching $492 million, and enterprise adoption of Devin has maintained a 50% month-over-month growth rate over the past six months.
Customers such as Goldman Sachs, Mercedes-Benz, Citigroup, Dell, Cisco, NASA, the U.S. Army, and the U.S. Navy have elevated its enterprise narrative beyond just demos.
A $26 billion valuation isn't for Devin, a perfect replacement for programmers already in place, but for the potential Cognition represents after its transformation: in the first industry to adopt AI Agents, it could become the new gateway for enterprise software engineering.
Future software development will likely not fully return to a time when human engineers code alone, nor will it immediately become fully automated by AI agents. A more predictable outcome is a hybrid system: humans in the IDE make strategic decisions while AI assists alongside; some tasks are delegated to cloud-based agents for asynchronous processing; and the code still undergoes testing, review, and merging, with human accountability remaining essential.
Cognition is betting on this middle ground.