Cognition launches SWE-2 coding model: 50.0% FrontierCode at 36% of Fable 5.1 cost
Cognition released SWE-2 on September 10, post-trained from Kimi K3. FrontierCode 1.1 Main scored 50.0%, within 1 point of Claude Fable 5.1 at 64% lower cost.
On September 10, Cognition released SWE-2, the successor to SWE-1.7 and the company's closest model to date to frontier-tier coding performance. SWE-2 is available immediately in Devin Desktop and Devin CLI, with rollout underway to Devin Web and Fusion.
On Cognition's in-house FrontierCode 1.1 Main benchmark, which measures whether AI-generated pull requests would actually be merged by a human maintainer, SWE-2 scored 50.0%, just 0.9 percentage points behind Claude Fable 5.1's 50.9%. Cognition claims SWE-2's run cost on this benchmark was 64% lower than Fable 5.1, roughly one-quarter that of OpenAI GPT-6 Astra.
The model is post-trained from Moonshot AI's open-source 2.8-trillion-parameter MoE model Kimi K3. Cognition's key technical move was training medium, high, and max reasoning effort levels in a single RL run, letting users trade more compute for harder tasks without changing models. SWE-2's medium setting reduces average turns by 58% and cost by 81% versus SWE-1.7, while moving the median first code edit from 48 steps down to 18.
On Terminal-Bench 4, a harder long-horizon agentic suite, SWE-2 scored only 27.3%, well behind Fable 5.1's 55.8% and GPT-6 Astra's 57.9%. Cognition disclosed this gap directly in its launch post and stated that the model's advantage concentrates on everyday coding tasks; long-horizon autonomous repo refactors remain the domain of frontier models.
On availability and pricing, Cognition has not published a standalone API price or model weights. SWE-2 is only available as part of the Devin product, with unlimited usage included for Pro and higher subscriptions starting at $20/month.