SWE-2 Now Available in Devin Desktop and CLI
WindsurfCognition launched SWE-2, its most advanced coding model yet, rolling it out first inside Devin Desktop and Devin CLI. Built by scaling reinforcement learning on top of the 2.8-trillion-parameter Kimi K3 base, SWE-2 scores 50.0% on Cognition's FrontierCode 1.1 benchmark, within a point of Claude Fable 5.1, while Cognition claims it costs up to 64% less. It also posts 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, though it has no standalone API or open weights, so it can only be used inside Devin's own products.
Featured Video
A video we selected to help illustrate this changelog
Key Takeaways
- On an independent KingBench 3 test, SWE-2 scored 83.75% (67 of 80 tasks), placing fifth on the reviewer's leaderboard, ahead of Fable 1.5, Kimi K3, and Claude Opus 5.
- The same review found SWE-2 particularly strong on longer, multi-step tasks, such as a data-generation-plus-fine-tuning job and a terminal-based movie tracker app, calling it better than Astra or Fable at those specific tasks.
- The reviewer's main complaint was that SWE-2 asks too many clarifying questions before starting work, slowing it down despite otherwise strong task performance.
- Despite beating some frontier models hands-on, SWE-2 still landed behind Qwen 3.8 Max, GPT-6 Astra, and GLM 5.3 on the same benchmark, tempering Cognition's "near-frontier" framing.
- On Hacker News, the 418-point discussion called out selective benchmarking, noting SWE-2's much weaker 27.3% score on the harder Terminal-Bench 4 compared to its strong 92.8% on Terminal-Bench 2.1.
- SWE-2 has no standalone API and no open weights, meaning it can only be used inside Devin Desktop, CLI, or Cloud, not through OpenRouter or a custom agent harness.
Sources & Mentions
3 external resources covering this update
A Near-Frontier Model Built for Cost
Cognition introduced SWE-2 on September 10, 2026, describing it as its closest model yet to the frontier. Rather than training a new model from scratch, Cognition post-trained SWE-2 from Moonshot AI's 2.8-trillion-parameter Kimi K3 base using reinforcement learning that explicitly accounts for task cost, not just correctness. The result, according to Cognition, scores in the same range as recent frontier models while costing as little as a third as much to run. SWE-2 is available starting immediately in Devin Desktop and Devin CLI, with Devin Web and Devin Fusion support planned to follow.
Benchmark Scores
On Cognition's own FrontierCode 1.1 Main benchmark, SWE-2 scores 50.0%, just below Claude Fable 5.1's 50.9% and ahead of Kimi K3 (44.2%), Grok 4.6 (48.0%), and GPT-5.6 Sol (47.5%), though behind GPT-6 Astra (53.3%). On DeepSWE 1.1 it scores 73.0%, and on Terminal-Bench 2.1 it scores 92.8%, both competitive with or ahead of most of the same comparison set. Cognition frames the headline number as SWE-2 matching Fable 5.1-level performance at 64% lower cost.
What Independent Testing Found
A widely discussed 418-point Hacker News thread pushed back on how selectively Cognition picked its benchmarks, pointing out that SWE-2's 92.8% on Terminal-Bench 2.1 does not carry over to the harder Terminal-Bench 4, where it manages only 27.3%. Commenters also flagged that SWE-2 has no standalone API and no open weights, so teams that want to call it through OpenRouter or plug it into their own agent harness or CI/CD pipeline cannot do so; it only runs inside Devin's own tools.
A separate hands-on video review using a custom eight-task "KingBench 3" benchmark scored SWE-2 at 83.75% (67 of 80 tasks), placing it fifth on the reviewer's leaderboard, ahead of Fable 1.5, Kimi K3, and Claude Opus 5, though behind Qwen 3.8 Max, GPT-6 Astra, and GLM 5.3. The reviewer singled out SWE-2's performance on longer, multi-step tasks, such as a data-generation-plus-fine-tuning job and a terminal-based movie tracker app, as a particular strength, while flagging a tendency to ask too many clarifying questions before starting work.
What This Means for Devin Users
For developers already inside the Devin ecosystem, SWE-2 gives Devin Desktop and CLI users a cheaper option that trades a small amount of peak benchmark performance for a meaningfully lower price, without needing to leave Cognition's own tooling. Its restriction to Devin's managed environment, however, means it is not a general-purpose model choice for anyone building or running agents outside of Devin itself.