AI
Cognition releases SWE-2 with a new approach to coding model efficiency
Cognition has released SWE-2 in Devin Desktop and CLI, with other Devin products following. The company reports a 50.0% score on FrontierCode 1.1 Main, approaching Fable 5.1 at 64% lower cost. Its central training change optimizes multiple reasoning effort levels in one reinforcement learning run using separate cost penalties. The disclosure offers useful engineering detail, but benchmark comparisons combine public results and internal evaluations using different agent frameworks.
Editorial summary · Publisher article
What matters
- SWE-2 is post-trained from the 2.8-trillion-parameter Kimi K3 model.
- Cognition reports that medium effort used 58% fewer turns and cost 81% less than SWE-1.7 on FrontierCode 1.1 Main while scoring higher.
- Training changes include more environments, stronger verifiers and an online draft model for speculative decoding.
Context & caveats
- Performance and behavioral findings are Cognition’s own claims; costs assume list pricing including public discounts.
- The supplied text omits mathematical expressions and some chart details, limiting scrutiny of the derivations and comparisons.
Why this made the edition
Detailed primary-source disclosure covers training methods, measured behavior, deployment and evaluation methodology. Missing mathematical notation and vendor-run comparisons limit verification but do not prevent a supported summary.
Source material