Cognition introduced SWE-2 on September 10, 2026, calling it its most advanced coding model so far. The model scores 50.0% on FrontierCode 1.1 Main, one point behind Fable 5.1 while costing 64% less. It is post-trained from Kimi K3, a 2.8T-parameter open model, making this Cognition's first RL run in the multi-trillion-parameter regime.

The central training innovation is a cost-penalized RL objective of the form R = S - lambda_e * C, where S is binary success, C is rollout cost in dollars and time, and lambda_e is tuned per effort level to the local slope of the base model's Pareto frontier. This lets Cognition train all reasoning-effort levels in a single RL run. The team proves in an appendix that only a linear cost penalty keeps the reward dependent solely on average cost and solve rate.

On benchmarks, SWE-2 reaches 73.0% on DeepSWE 1.1, 92.8% on Terminal-Bench 2.1, and 27.3% on Terminal-Bench 4. It beats SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5.1 at a fraction of their price, and sits a few points behind GPT-6 Astra at about a quarter of the cost. Its RL added 5 to 6 points over the Kimi K3 base on many benchmarks.

Efficiency gains are large. SWE-2 medium outperforms SWE-1.7 on FrontierCode 1.1 Main while taking 58% fewer turns and costing 81% less. It makes its first real edit after a median of 18 steps versus 48 for SWE-1.7, which Cognition attributes to more focused codebase exploration. Behaviorally, the team reports better end-to-end test writing, more resourcefulness when a path is blocked, and stronger verification habits.

On the infrastructure side, Cognition details a length-weighted reward baseline that cuts gradient variance and keeps inference-training KL low, a prefill delayer that raised throughput 10 to 20%, and an online-trained DSpark speculative decoding draft model with 15% longer accept lengths. NVFP4 and FP8 kernels with quantization-aware training hold memory use down on a base model nearly three times larger than before.

Data work tripled the number of RL environments, added instruction-following overlays, and used previous SWE-2 checkpoints to iteratively harden verifiers against reward hacking. A trustworthiness evaluation covering propaganda, censorship, and context-dependent vulnerability found a 98.0% overall pass rate on politically sensitive China-related questions and no statistically significant framing effects on coding vulnerability.

SWE-2 is available now in Devin Desktop and the Devin CLI, with rollout to Devin Web and Fusion underway.