Cognition, the company behind an AI software engineer called Devin, raised over $1 billion at a $26 billion valuation this week. The round was led by Lux Capital, General Catalyst, and 8VC, and the company has now raised over $2.5 billion in total.
The valuation jump is the headline, but the revenue numbers are what make it credible. Cognition's annual revenue run-rate grew from $37 million in May 2025 to $492 million today. That is a 13-fold increase in twelve months, and the company says enterprise usage has grown more than tenfold since January 2026 alone.
Devin is not an autocomplete tool that suggests lines of code as a developer types. It works differently: you assign it a task in plain language, it creates its own working environment, reads the relevant codebase, plans a sequence of steps, runs tests, fixes its own errors, and submits the result for a human to review. The human does not watch every step; they review the output at the end. Two years ago, 34% of Devin's submitted work was accepted. Today, that figure is 67%, double the rate from a year ago.
The customer list tells you this has moved past the technology industry. Goldman Sachs, Citi, Santander, Mercedes-Benz, Dell, Itaú, Fiserv, the U.S. Army, and the U.S. Navy are all using it. Fiserv, which provides core banking technology to thousands of financial institutions, announced a partnership with Cognition just days ago to use Devin for modernising banking infrastructure. Itaú, Latin America's largest bank, uses it to automatically resolve 70% of its security vulnerabilities. Mercedes-Benz used it to compress an eight-month technology modernisation project into eight days.
The most striking data point in the announcement is internal. Cognition says that 89% of all code its own engineers commit is now written by Devin, up from just 13% in December 2025. Five months. That is either the most honest advertisement a software company has ever published, or a preview of what other organisations will be looking at within the next few years.
Devin performs best on well-defined, repeatable tasks: migrating old code to a newer format, fixing security vulnerabilities flagged by scanning tools, writing automated tests, and completing standard features that follow existing patterns in a codebase. It is less suited to exploratory work, tasks with shifting requirements, or complex architectural decisions. Those still require experienced humans.
The competitive picture is worth understanding. GitHub Copilot, backed by Microsoft, has 4.7 million paying users and works like an in-editor suggestion tool. Cursor, valued at $29 billion, earns roughly $2 billion in annual revenue and also works primarily inside a code editor. Anthropic's Claude Code is growing sharply and leads developer satisfaction surveys. Devin sits in a different category: it takes tasks fully off a person's plate rather than assisting while they work. Several enterprise teams use Devin alongside Cursor, treating them as complementary rather than competing.
The broader market context: enterprise spending on AI coding tools is estimated at roughly $10 billion annually as of mid-2026, and the autonomous agent segment within that is growing at the fastest rate of any sub-category. Gartner found that 90% of engineering leaders report productivity improvements from these tools, with an average gain of 19% across organisations. The gap between companies that have adopted this and those that have not is already showing up in how fast they can modernise, ship, and maintain software.
For business operators outside software: your IT teams, your vendors, and your systems integrators are already using or evaluating tools like this. The question worth asking internally is not whether AI is writing code somewhere in your supply chain, but whether your organisation has visibility into where, and whether the governance and review processes are in place to manage it.