Evaluating Cognition's SWE-2 Coding Model
AICodeKinggo watch the original →
the gist
Cognition's SWE-2 model demonstrates high consistency and strong performance on complex, multi-step agentic tasks, outperforming DeepSeek V4.1 Flash on KingBench 3 with an 83.75% score.
Performance on KingBench 3
SWE-2 achieved a total score of 67 out of 80 (83.75%) across eight varied coding tasks, placing it fifth on the KingBench 3 leaderboard. The model demonstrated high consistency, scoring no lower than 7 out of 10 on any individual task. It outperformed DeepSeek V4.1 Flash, which scored 65 out of 80 (81.25%). SWE-2 showed particular strength in multi-stage workflows, such as generating a dataset, fine-tuning a Gemma 2B model, and building a local web interface, where it achieved a perfect 10 out of 10.
Operational Characteristics and Limitations
While the model excels at complex, connected tasks, it exhibits a tendency to request excessive clarification before beginning work. This behavior creates unnecessary back-and-forth for routine implementation decisions. The author suggests that users mitigate this by providing explicit system instructions to limit clarifications to one per request and to make reasonable assumptions for standard architectural or layout choices. Despite this, the model remains a highly capable option for long-form agentic coding, showing a clear performance improvement over its base model, Kimi K3.
Pricing and Availability
Cognition currently includes access to SWE-2 within the $20-per-month Devin Pro plan. This offer is valid through October 10, 2026, making it a cost-effective choice for developers looking to integrate agentic coding tools into their existing workflows.