Cognition, the maker of the autonomous coding agent Devin, introduced its newest coding model, SWE-2, this week, positioning it as a cost-performance breakthrough rather than a pure capability leap. The company says SWE-2 scores 50.0% on the FrontierCode 1.1 Main benchmark — within a single point of Anthropic's Claude Fable 5.1 — while costing roughly 64% less to run.

Built on an open-weight base

Unlike Cognition's earlier SWE-1.7 model, SWE-2 is post-trained on top of Moonshot AI's open-weight Kimi K3, a 2.8-trillion-parameter mixture-of-experts model, rather than a proprietary base. Cognition says this is the first time reinforcement learning has been scaled to a base model of that size for a coding-specific product. The RL recipe reportedly trains all of the model's reasoning-effort tiers — from fast, cheap responses to slower, more thorough ones — in a single training run, which the company says is what let it push the whole cost-to-performance curve rather than trading one for the other.

On efficiency, Cognition reports that SWE-2's medium setting outperforms SWE-1.7 while needing 58% fewer turns to finish a task and costing 81% less to do so, with the average number of steps per coding run dropping from 127 under SWE-1.7 to 53 under SWE-2 medium.

Rolling out across Devin

SWE-2 became available immediately in Devin Desktop and Devin CLI, with support for Devin Web and Cognition's Fusion product rolling out afterward. Devin is Cognition's flagship product, an agent designed to work through real software engineering tasks — reading a codebase, planning changes, writing and testing code, and fixing its own mistakes — largely without step-by-step human supervision.

Why it matters

The release lands one week after Cognition closed a Series E round valuing the company at $48 billion, and it doubles as an answer to a question that funding round raised: whether Cognition can keep pace with frontier labs on raw model quality without matching their training budgets. By building on an openly available base model instead of training a frontier model from scratch, Cognition is betting that being close to frontier-level coding performance at a fraction of the cost matters more to enterprise engineering teams than being outright best. That bet puts pressure on rivals like GitHub Copilot, Cursor and Anthropic's own Claude Code to compete on price as well as capability in the fast-growing autonomous coding-agent category.