Cognition launches new SWE-2 model, Rivaling Fable 5.1 and GPT-Astra

The cost of using advanced AI coding assistants drops by up to 64%, making high-performance capabilities more accessible.

Cognition has launched SWE-2, their latest and most advanced coding model, positioned as a significant advancement in AI-powered software development. SWE-2 achieves 50.0% on the FrontierCode 1.1 Main benchmark, placing it within one point of Fable 5.1, but at a 64% lower cost. This performance was achieved by scaling Reinforcement Learning (RL) to the multi-trillion-parameter regime, incorporating an RL algorithm that trains all reasoning-effort levels in a single run. The model is post-trained from Kimi K3, a 2.8T-parameter model, and demonstrates notable improvements in efficiency and intelligence. SWE-2 is designed to reduce unnecessary exploration, write more comprehensive tests, and exhibit greater resourcefulness when encountering obstacles. It is available immediately on Devin Desktop, CLI, Web, and Fusion platforms.

Cognition's SWE-2 introduces a novel RL training approach that optimizes the entire cost-performance frontier simultaneously, rather than optimizing individual points. This method, which applies a linear cost penalty tuned to the base model's Pareto frontier slope, directly integrates user cost considerations into the training process. This principled approach allows for significant gains in both capability and efficiency, making advanced coding assistance more economically viable for a wider range of users.

The introduction of SWE-2 signals a strategic shift towards more cost-efficient AI models that do not compromise on performance. By achieving near-top-tier results at a fraction of the cost, Cognition is challenging established players and redefining the economic landscape of AI coding assistants. This development will likely pressure competitors to not only match performance but also to address cost-effectiveness, potentially leading to a broader market adoption of sophisticated AI development tools.