Devin hits $492M run-rate as Cognition bets on independent agents
Original: More Devins in More Places View original →
The strongest signal in Cognition’s new round is not only the $1B-plus check. It is the claim that Devin has reached $492M in run-rate revenue while enterprise usage has grown more than 10x since the start of 2026. Cognition said on May 27, 2026 that its Series D values the company at $26B.
The round was led by Lux Capital, General Catalyst, and 8VC. Existing backers including Founders Fund, Elad Gil, Alpha Wave, Definition Capital, Positive Sum, Avenir, Vitruvian, Bain Capital Ventures, Conversion Capital, 137 Ventures, Soma Capital, and Omri Casspi also participated. New investors include Ribbit Capital, Atreides, and Layer Global.
Cognition launched Devin two years ago as an AI software engineer. The company now says cloud agents are the fastest-growing way to create software, and it names customers including Citi, Mercedes-Benz, Goldman Sachs, Elevance, Dell, Santander, the U.S. Army, and the U.S. Navy. It also cites specific outcomes: Mercedes-Benz cut an eight-month legacy modernization project to eight days, while Itaú automatically fixes 70% of security vulnerabilities with Devin.
The strategic bet is in Cognition’s description of itself as an “Independent Agent Lab.” Rather than tying Devin to one foundation model, the company says it works with model labs, evaluates performance across 100-plus categories of software engineering tasks, and routes work based on price and performance. That matters because token usage is growing quickly, and engineering teams are increasingly forced to manage agent spend as carefully as output quality.
The valuation now prices Devin as more than a coding assistant. It prices Cognition as a workflow layer that can sit between enterprises and multiple model providers. The next proof point will be whether customers keep measuring Devin by shipped pull requests, vulnerability fixes, and modernization timelines after the novelty of autonomous coding has worn off.
Related Articles
The HN debate centered on measurement: equal pass rates do not mean equal agent cost or navigation behavior.
AI coding is shifting from picking one assistant to orchestrating several agents. Omnigent is an open-source meta-harness with shared sessions, guardrails, and human-in-the-loop workflows.
OpenAI says SWE-Bench Pro no longer reliably measures frontier coding capability after finding 30% of its public tasks broken. The cited issues include hidden requirements, contradictory instructions, strict tests and incomplete grading criteria.