Cognition's SWE-2 nears Fable 5.1 on FrontierCode at 64 percent less cost
single source · 1 articles · release · confidence: medium · first seen 2026-09-12 23:56 UTC
Cognition, the company behind the Devin coding agent, has released SWE-2, a coding model post-trained with reinforcement learning (reward-based training) on Kimi K3, Moonshot AI's 2.8T-parameter open model. On FrontierCode 1.1 Main, SWE-2 scores 50.0%, which Cognition says is within 1 point of Fable 5.1 while costing 64% less. The company describes SWE-2 as its most capable coding model to date. The 64% lower cost is relative to Fable 5.1, as reported by Cognition. The release does not specify API availability or further benchmark details. No independent evaluation is cited in the release.
What this means for you
If you currently use Fable 5.1 for coding tasks, SWE-2 is worth benchmarking: Cognition reports near-identical FrontierCode performance at 64% lower cost. Check your provider for availability and run your own tests before switching.
Key facts
- ·Cognition released SWE-2, its most capable coding model to date. source
- ·SWE-2 is post-trained with reinforcement learning from Kimi K3, Moonshot AI's 2.8T-parameter open model. source
- ·SWE-2 scores 50.0% on FrontierCode 1.1 Main, which Cognition says is within 1 point of Fable 5.1. source
- ·Cognition reports SWE-2 costs 64% less than Fable 5.1. source
- ·Cognition is the company behind the Devin coding agent. source
What the sources say
- MarkTechPost — Reports the SWE-2 release, its FrontierCode score, and the claimed 64% cost reduction.
Sources
The original reporting. Follow these — they did the work.