Cognition introduced SWE-2, its latest coding AI model, on September 10. SWE-2 achieves a 50.0% score on the FrontierCode 1.1 Main benchmark, placing it within one point of Fable 5.1. The model delivers this performance while being 64% cheaper than comparable offerings, marking a significant advance in cost-efficiency and capability, according to cognition.com.
The SWE-2 model builds on the training infrastructure of its predecessor, SWE-1.7, scaling reinforcement learning (RL) to the multi-trillion-parameter level for the first time. A new RL algorithm trains all reasoning-effort levels simultaneously, pushing the cost-performance frontier forward. SWE-2 is post-trained from the Kimi K3 model, a 2.8 trillion-parameter AI that had already undergone extensive RL for agentic coding, enabling SWE-2 to add 5 to 6 points on many benchmarks beyond K3’s baseline.
SWE-2 outperforms previous Cognition models like SWE-1.7 and Grok 4.6 in both score and cost. It matches the performance of GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price and comes close to GPT-6 Astra’s performance at about a quarter of the cost. This positions SWE-2 as a competitive option in the AI coding model market, where balancing cost and capability is critical for adoption.
Cognition’s FrontierCode leaderboard tracks model rankings, with SWE-2 now near the top. The company’s blog post dated September 10 details these advancements and the model’s benchmark results, highlighting SWE-2’s role in advancing AI coding efficiency and affordability.