Cognition's SWE-2

SWE-2 is Cognition's latest coding model, designed to push the Pareto frontier of capability and cost. It scores 50.0% on the FrontierCode 1.1 Main benchmark, within one point of Fable 5.1 while being 64% cheaper. The model is built on the SWE-1.7 training infrastructure and introduces a new RL algorithm that trains all reasoning-effort levels in a single run, advancing the cost-performance frontier.

SWE-2 stands out for its strong engineering judgment and verification discipline. It explores codebases thoroughly before making edits, writes end-to-end tests that catch regressions, and re-derives conclusions when challenged rather than re-asserting. It also handles complex tasks by planning more, exploring more of the codebase, and managing uncertainties through complex verification. SWE-2 is available in Devin Desktop, CLI, and Web, and integrates with Fusion.

This model is ideal for developers and teams who need a coding agent that can handle complex, multi-step tasks with high reliability. It's particularly useful for large codebases where thorough exploration and verification are critical. With its cost efficiency, SWE-2 offers a compelling option for organizations looking to scale AI-assisted coding without breaking the bank.