Cognition's Devin and the AI coding race

Cognition now has more than 200 employees, up from 40 last summer, after buying Windsurf. That growth says a lot about where the AI coding business is headed: fast, expensive, and increasingly enterprise-focused. The company’s flagship product, Devin, is pitched as an AI software engineer that can handle real tasks from start to finish.

That pitch matters because the market around it is getting crowded in a hurry. Cursor, Claude Code, and OpenAI Codex are all chasing the same prize: software that can do more of the boring, repetitive parts of engineering without sitting on a human’s shoulder the whole time.

The bet behind Devin

Cognition was founded in late 2023 around a simple question: can AI do more than answer prompts and actually complete engineering work? CEO Scott Wu, a former competitive programming prodigy with three gold medals from the International Olympiad in Informatics, thinks the answer is yes. He frames Devin as a tool that can turn ideas into software while leaving people to decide what gets built.

That framing is important. A lot of AI coding tools still feel like very smart autocomplete. Devin is supposed to be closer to a junior engineer that can browse the web, inspect a codebase, make a plan, try one approach, then switch tactics when the first attempt fails. That is a much bigger claim than “helps you code faster.”

The company’s early demo story makes the point well. One of Cognition’s founders asked Devin to get a refund for a rescheduled flight. Devin found the airline’s site, read the terms, chatted with a bot, then escalated to a human when the automated route stalled. It is a small example, but it shows the kind of multi-step work Cognition wants Devin to handle inside software teams.

• Founded in late 2023
• Raised $400 million in a round led by Founders Fund
• Valuation above $10 billion
• Expanded from 40 employees to 200+ after the Windsurf deal

Why the office looks like a factory floor for code

Cognition’s culture sounds closer to a high-pressure lab than a normal startup. The company moved from an Atherton mansion into a 25,000-square-foot headquarters in San Francisco’s South Park, and the office setup reflects the pace Wu expects. There is an in-house barista, private chefs, a gym stocked with Devin-branded creatine, and even an underground speakeasy with poker and video games.

That environment is part branding, part retention strategy, and part signal to the market. Cognition wants employees, investors, and customers to believe the company is obsessed with shipping. Wu has described the workplace as an “extreme performance culture,” and staff have said they often work late into the night and through weekends.

“We’re working toward giving everyone their own buddy, Devin, who can go and build out their ideas for them into real software and products.” — Scott Wu

The culture also has a practical side. AI coding products improve quickly when teams push on edge cases, collect failure modes, and keep iterating. That is easier to do when the company is organized around one product and one goal. It is harder to do when a startup is trying to be a general-purpose AI platform with a dozen bets running at once.

Cognition’s new office setup is also a sharp break from its earlier days in a $10.5 million mansion in Atherton. Back then, the team turned bedrooms into workspaces, converted a wine cellar into a nap room, and kept a private chef on site. The mansion era is over, but the work intensity clearly is not.

How Cognition compares with Cursor, Anthropic, and OpenAI

The AI coding race is no longer a tiny club. Cognition is fighting for attention against companies with more users, more capital, and broader distribution. Cursor has become the default tool for many developers who want AI assistance inside their editor. Anthropic’s Claude Code has surged with developers who want a command-line style assistant. OpenAI’s Codex brings the weight of a giant platform and a huge installed base.

What makes Cognition interesting is that it is aiming higher than code completion or chat-based help. It wants to own the workflow of software creation itself. That is a harder product to build, but it could be more valuable if it works across enterprise teams that need agents to handle real tasks inside large codebases.

• Cursor has reached more than $2 billion in annualized revenue
• Claude Code has also passed $2 billion in annualized revenue
• Cognition says enterprise usage grew about 80-fold over the past year
• Customers include Goldman Sachs, Citi, and NASA

Those numbers matter because they show where the money is. Consumer hype gets the headlines, but enterprise buyers pay for reliability, security, and integration into existing workflows. Cognition says that is where it has found its strongest footing. It did not share revenue, and it did not say how many consumers use Devin, which suggests the enterprise side is doing more of the heavy lifting.

The competition is also forcing a sharper product question: what counts as an AI software engineer? If a tool can write tests, debug, and open pull requests, that is useful. If it can plan an entire feature, inspect dependencies, and recover from errors without constant supervision, that is a different class of product. Cognition is trying to prove it belongs in the second category.

What this says about the next phase of software work

The bigger story here is not that engineers disappear. It is that the job keeps splitting in two. The repetitive parts of coding get automated faster, while the human parts become more valuable: deciding what to build, checking quality, and making tradeoffs when the system gets something almost right.

Wu’s argument is that software abundance follows from that split. If building gets cheaper and faster, more teams will ship more software. That is the bet behind Devin, and it is also why Cognition is willing to spend heavily on talent, office space, and product iteration while the market is still sorting out what AI coding should become.

My read: the next 12 months will tell us whether Devin is a real enterprise workflow tool or just the most polished demo in the category. If Cognition can keep growing inside large companies while Cursor and Claude Code keep pulling in developers, the AI coding market may split into editor helpers, autonomous agents, and a few expensive systems that do both. Which bucket Devin lands in will decide whether Cognition becomes a name developers use every day or just a very loud chapter in the race to automate software work.