Cognition's SWE-2 is a coding model built on Kimi K3
2026/09/15
Cognition announced SWE-2, its new coding model, on 10 September 2026[1]. It is available from that day in Devin Desktop and the CLI, and is rolling out to Devin Web and Fusion[1]. The announcement gives no API price[1].
SWE-2 is post-trained from Moonshot's Kimi K3, a 2.8T-parameter model that had already been through extensive reinforcement learning for agentic coding[1,2].
Cognition says it scaled reinforcement learning to the multi-trillion-parameter range, with an algorithm that trains every reasoning-effort level, medium, high and max, in a single run[1].
On FrontierCode 1.1 Main, SWE-2 scores 50.0 percent, ahead of Grok 4.6 at 48.0 and the earlier SWE-1.7 at 42.0[1].
Cognition puts it within one point of Fable 5.1 at 64 percent lower cost, and a few points behind GPT-6 Astra at a quarter of the cost[1].
Its other reported scores are 73.0 percent on DeepSWE 1.1, 92.8 on Terminal-Bench 2.1 and 27.3 on Terminal-Bench 4[1].
The appendix sets out how the comparison was made[1]. Public results are used where they exist, and otherwise Cognition ran the model itself[1].
In that case each model runs in the harness it was mainly built for: Claude Code for Anthropic models, Codex for OpenAI, Grok Build for xAI and Devin CLI for open-weight models[1].
Costs assume list prices including public discounts, and each score is the best across reasoning-effort settings[1]. With a different harness for each model, the table does not compare like with like.
