Cognition Raises $2 Billion at a $48 Billion Valuation, Closing In on Cursor

Cognition has pulled in more than $2 billion in a Series E and now sits at a $48 billion valuation. That is a sharp climb from the spring, when the company was priced at $26 billion. The startup is still barely out of the gate — it was founded in 2024 — but the market is already treating it like one of the two pillars of AI-assisted software development.

The round was led by Andreessen Horowitz and Accel, both new investors in Cognition. Founders Fund, General Catalyst and Avenir came back, and so did a long list of heavy hitters: Benchmark, Bessemer, Kleiner Perkins, Greylock, Lightspeed, Altimeter, Bond Capital, Meritech, T. Rowe Price, Lux Capital, 8VC, DST, Bain Capital Ventures, Ribbit, Battery, Stripes, Alpha Wave and Alkeon. Cognition says more than 30 funds joined in total. That’s a crowded cap table for a company that still talks like it’s only getting started.

The money is landing on top of a business that says its annualized run rate is close to $900 million. In the spring, that figure was $492 million. That’s the sort of jump that makes investors reach for the next term sheet before the ink is dry on the last one. It also helps explain why Cognition can raise at such a pace: the company is not selling hope alone.

Devin, its AI coding agent, is pitched as something more ambitious than a slick autocomplete. Cognition wants engineers to behave more like architects, handing execution to swarms of agents. The product now includes Auto-Triage for incident investigation, Security Swarm for finding and triaging vulnerabilities, and Automations that kick off tasks from Slack, GitHub, Linear and other systems. The company’s bigger pitch is “self-driving software,” where agents act on their own and compute shifts toward the work that matters most.

Cognition is also trying to keep itself model-agnostic. Devin currently uses both OpenAI’s GPT-6 Astra and Anthropic’s Claude Fable 5.1, and Cognition says Astra lands within 0.4 points of Fable 5 on its FrontierCode 1.1 benchmark while costing 64 percent less. That kind of gap is the whole argument for a multi-model shop. Meanwhile, the company is building its own SWE family for software engineering, with SWE-1.7 introduced in the summer and based on Kimi K2.7 before Cognition pushed it further with its own reinforcement learning. The office map has widened too, from San Francisco, New York and Austin to Washington, D.C., Tokyo, Singapore, London, São Paulo and Madrid.