Cognition has dropped SWE-2, its most capable coding agent yet, and the numbers suggest a significant shift in the cost-performance curve for automated software engineering. The model scores 50.0% on the FrontierCode 1.1 Main benchmark, placing it within a single point of Fable 5.1 but at 64% lower cost. This isn’t just a marginal upgrade; Cognition claims SWE-2 operates at roughly a quarter of the cost of GPT-6 Astra while staying competitive on most metrics. The secret sauce lies in its training pipeline, which post-trains on Moonshot AI’s massive 2.8T-parameter Kimi K3 model using a novel reinforcement learning approach.
The technical leap here is efficiency through focused exploration. Previous iterations like SWE-1.7 tended to over-explore simple tasks, burning tokens and time. SWE-2 changes this behavior dramatically. On FrontierCode, the medium effort level takes 58% fewer turns and costs 81% less than its predecessor. Mean steps per run drop from 127 to just 53 for the medium setting. Cognition achieved this by training all three reasoning-effort levels in a single RL run, using pareto-informed cost penalties that force the model to optimize both success rate and inference cost simultaneously. It’s a clever mathematical trick that moves the entire cost-and-performance frontier up at once.
However, developers shouldn’t expect to download weights or call an API. SWE-2 remains locked inside the Devin ecosystem, available via Desktop and CLI today with Web and Fusion versions coming soon. While the performance is impressive, there are clear weak spots. SWE-2 trails competitors by roughly 30 points on Terminal-Bench 4, suggesting it still struggles with complex terminal interactions compared to high-end rivals. Additionally, all rival benchmark numbers come from Cognition’s own evaluations, so independent verification remains pending.
This release cements Cognition’s strategy of vertical integration. By controlling the model, the agent, and the interface, they can optimize the full stack in ways open-weight providers cannot. But it also means you’re buying into their walled garden. With free access for paid tiers through October 2026, the barrier to entry is low, but the lock-in is real. The industry now faces a choice: build on proprietary agents that deliver immediate ROI, or wait for open models to catch up to this level of tuned efficiency.
