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

Cognition has unveiled SWE-2, a groundbreaking coding model that matches the capabilities of Fable 5.1 and GPT-Astra while being significantly more cost-effective. This new model represents the pinnacle of scaling reinforcement learning to multi-trillion-parameter regimes, a milestone previously unattained.

Key differentiators of SWE-2 include its ability to train all reasoning-effort levels concurrently, a leap forward in managing cost and performance. When benchmarked against FrontierCode 1.1 Main, SWE-2 delivers a 50.0% score, only 1 point shy of Fable 5.1's 51.0%, and does so at a 64% lower cost. The model's efficiency is further highlighted by its ability to make its first edit after only 18 steps on average, compared to 48 steps for its predecessor, SWE-1.7.

SWE-2's superiority extends to complex tasks as well, where its planning capabilities and uncertainty management set it apart from GPT-5.6 Sol and Fable 5/5.1. Remarkably, SWE-2 achieves these high benchmarks at a mere quarter of the cost of GPT-6 Astra.

Post-training enhancements, including Pareto-informed cost penalties in reinforcement learning, have been instrumental in shaping SWE-2's unique characteristics. These advancements allow for a more nuanced cost-performance tradeoff, aligning the model's intelligence with its economic efficiency. The result is a model that not only rivals industry heavyweights like Fable 5.1 and GPT-Astra but also does so at a fraction of their price, marking a significant leap in the coding AI landscape.