SWE-2
A coding model from Cognition that, per official claims, delivers near-frontier capability at 64% lower cost, available in Devin Desktop and CLI.
About SWE-2
SWE-2 is a coding model released by Cognition, which the company describes as its most advanced coding model to date. It officially scores 50.0% on FrontierCode 1.1 Main, less than one point behind Fable 5.1, while costing 64% less. The model is available starting today in Devin Desktop and the Devin CLI, and is rolling out gradually on Devin Web. It is primarily aimed at developers completing real-world software engineering tasks through the Devin family of AI coding tools.
Best for
Best use cases
Use a stronger coding model in Devin Desktop or the CLI for everyday development tasks
Choose a coding model that approaches frontier-level performance while keeping inference costs in check
Compare before choosing
According to the official blog: on FrontierCode 1.1 Main and DeepSWE 1.1, SWE-2 outperforms SWE-1.7 and Grok 4.6 in both score and cost; matches GPT-5.6 Sol and Fable 5/5.1 at a far lower price; and achieves results close to GPT-6 Astra at roughly a quarter of the cost.
Features
50.0% on FrontierCode 1.1 Main
According to the official announcement, it scored 50.0% on the FrontierCode 1.1 Main benchmark, less than one point behind Fable 5.1.
Near-frontier capability at lower cost
According to the official announcement, its capability is less than one point behind Fable 5.1, while costing 64% less.
Upgraded training approach
According to the official announcement, the number of RL environments was tripled, an instruction-following overlay was added, and checkpoints from before SWE-2 were used to build a flywheel of iterative reinforcement validators.
Pros
- Officially scores 50.0% on FrontierCode 1.1 Main, less than one point behind Fable 5.1, at 64% lower cost
- Officially matches the scores of GPT-5.6 Sol and Fable 5/5.1 on FrontierCode 1.1 Main and DeepSWE 1.1 at lower cost
- Available starting today in Devin Desktop and the CLI, with a gradual rollout to Devin Web