Cognition released SWE-2 on Sept. 10, saying its new software-engineering model performs alongside recent frontier coding systems at up to 70% lower cost. The company said SWE-2 is available immediately in Devin Desktop and CLI.
In a post on X and a blog post, Cognition said SWE-2 scored 50.0% on FrontierCode 1.1 Main, compared with 50.9% for Fable 5.1. Cognition said its model was 64% cheaper than Fable 5.1 on that measure.
For users of Devin, Cognition’s autonomous software engineer, the release adds a new model option in the company’s desktop and command-line products. Cognition said it is also rolling SWE-2 out on Devin Web and Fusion.
Cognition listed SWE-2 at 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1. On Terminal-Bench 4, it listed SWE-2 at 27.3%, versus 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra.
The company said SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter model. Cognition said it scaled its reinforcement-learning training recipe to the multi-trillion-parameter range for the first time, using infrastructure and a recipe developed for SWE-1.7.
SWE-2 is Cognition’s first model with effort levels, according to the blog post. The company said one reinforcement-learning run trains all of those levels together. On FrontierCode 1.1 Main, Cognition said SWE-2 medium scored higher than SWE-1.7 while taking 58% fewer turns and costing 81% less on average. It reported mean steps per run of 53 for SWE-2 medium, compared with 127 for SWE-1.7.
Cognition said its benchmark costs use list pricing that includes public discounts. It said each reported score reflects the best result across reasoning-effort settings using the harness for which the model was primarily developed.
Cognition said SWE-2 will be free for Pro, Max and Teams subscribers for the next month. The launch follows Cognition’s funding announcement, which Go Big News reported on Sept. 9.
