SWE-2
Cognition's coding model for Devin, post-trained from Kimi K3 — within a point of Fable 5.1 on FrontierCode, on its own figures.
About this launch
SWE-2 is Cognition's software-engineering model, released on 10 September 2026 and available from launch in Devin Desktop and the Devin CLI, with Devin Web and Fusion following. It is post-trained with reinforcement learning from Moonshot AI's open-weights Kimi K3, building on the training recipe of Cognition's earlier SWE-1.7, and its training objective counts the cost of each attempt, a mix of inference spend and time, as well as whether it succeeded. It offers three reasoning-effort levels (medium, high and max), which Cognition trained in a single run. Cognition reports 50.0% on FrontierCode 1.1 Main, within a point of Claude Fable 5.1 while being 64% cheaper, 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1; on Terminal-Bench 4 it reports 27.3%, well behind Fable 5.1 at 55.8% and GPT-6 Astra at 57.9%. These are vendor figures, each the best of SWE-2's effort settings, and the cost comparison assumes list prices with public discounts. Cognition's announcement states no price, context window or licence, so none is recorded; it is offered inside Devin rather than announced as a standalone API.
Discussion
Questions & answers about SWE-2.
No questions yet. Be the first to ask about SWE-2.
More on TrustList
Everything here links back to the same verified catalogue. Pick your next stop.
- All launchesToday's board, this week's and the all-time archive.
- SWE-2 on TrustListReviews, pricing, features and alternatives.
- CompaniesAgencies, consultancies and IT service providers, ranked by verified reviews.
- ProductsSoftware and SaaS with pricing, features, integrations and alternatives.
- ArticlesGuides, comparisons and research from the editorial desk and the community.
- AwardsAnnual recognition decided by verified reviews and an independent jury.
- AI ModelsBenchmark scores and community ratings for every major model.
- RequestsBuyers describe what they need; vendors respond directly.
- PeopleReviewers, authors and makers with public profiles.
- ComparePut up to four listings side by side before you shortlist.