Cognition announced the release of its SWE‑2 coding model, which it says scores 50.0% on the FrontierCode 1.1 Main benchmark and matches Fable 5.1 at a claimed 64% lower cost.
In a tweet on September 10, Cognition said SWE‑2 is post‑trained on Kimi‑K3 and that the model “achieves a score of 50.0%” on FrontierCode, beating its predecessor SWE‑1.7 as well as Grok 4.6 and GPT 5.6 Sol while matching Fable 5.1 at the lower cost.
Community user @thefounderspack reported the same 50.0% score, noting it is within one point of Fable 5.1 and that the model was post‑trained from Kimi K3, a 2.8‑trillion‑parameter base. The post added that SWE‑2 is currently available only in Devin Desktop and CLI and that no public API has been released.
Another community member, @ayyazdev, said the model was trained with all effort levels in a single run using a cost penalty matched to the local Pareto slope. According to the post, SWE‑2 makes its first real edit at a median step of 18, and the earlier SWE‑1.7 model scored 48 on the same benchmark.
A third community post by @BTA_labs listed additional numbers from Terminal‑Bench 4, showing SWE‑2 at 27.3%, Fable at 55.8% and Astra at 57.9%. The tweet reiterated the Kimi‑K3 post‑training and confirmed the model’s presence in Devin Desktop and CLI without a standalone API.
Sources
Fan coverage from Devin Central — not an official Cognition announcement. Devin is a trademark of Cognition.