Cognition has announced the release of SWE-2, marking a significant advancement in AI-powered coding assistance that prioritizes both performance and cost efficiency. The new model achieves 50.0% accuracy on the challenging FrontierCode 1.1 Main benchmark while operating at 64% lower cost than comparable systems, positioning it within one percentage point of the leading Fable 5.1 model.

The development represents a breakthrough in reinforcement learning scaling, as SWE-2 becomes the first model to successfully apply RL techniques to multi-trillion-parameter architectures. Built upon the SWE-1.7 training infrastructure, the system uses Kimi K3 as its foundation—a 2.8 trillion parameter model that had already undergone extensive RL training for coding applications.

A key innovation lies in SWE-2's unified training approach, which simultaneously optimizes multiple reasoning-effort levels in a single RL run. This methodology contrasts with traditional approaches that require separate training cycles for different performance tiers. The result is a model that can offer various capability levels at corresponding price points, effectively optimizing the entire cost-performance frontier rather than just peak performance.

Benchmark evaluations demonstrate SWE-2's competitive positioning across multiple assessment frameworks. On DeepSWE 1.1, the model achieves 73.0%, significantly outperforming its predecessor SWE-1.7 (37.7%) and competing effectively against Grok 4.6 (67.5%). The system matches the performance of GPT-5.6 Sol and Fable 5/5.1 while operating at substantially reduced costs, and approaches GPT-6 Astra's capabilities at approximately one-quarter the expense.

Behavioral improvements address previous user feedback about over-exploration tendencies in earlier versions. SWE-2 demonstrates more focused codebase analysis, beginning actual implementation after a median of 18 steps compared to SWE-1.7's 48 steps. This efficiency translates to practical benefits: 58% fewer interaction turns and 81% lower average costs on FrontierCode tasks.

The technical foundation relies on sophisticated cost-penalized reward functions that align training objectives with real-world usage patterns. Cognition's approach employs linear cost penalties calibrated to match the slope of the base model's performance curve at each effort level. This ensures that training improvements benefit the entire capability spectrum rather than optimizing solely for peak performance scenarios.

SWE-2 exhibits enhanced capabilities across multiple dimensions of software development. The model shows improved test coverage abilities, writing more comprehensive end-to-end tests that better catch regressions and edge cases. It demonstrates increased resourcefulness when facing obstacles, finding alternative approaches when primary solutions are blocked. Additionally, the system shows stronger verification discipline, re-deriving conclusions when challenged rather than simply reasserting previous statements.

The model offers differentiated behavior across effort levels, with SWE-2 medium providing quick action for simple and intermediate tasks, while SWE-2 high and max variants excel on complex problems through more extensive planning and codebase exploration. This tiered approach allows users to select appropriate capability levels based on task complexity and budget constraints.

Technical innovations extend to the training infrastructure, including improved RL rollout serving with enhanced scheduling and online draft model training. The system incorporates NVFP4/FP8 kernels and quantization-aware training to reduce memory usage while maintaining lower train-inference mismatch compared to SWE-1.7, despite using a base model with nearly three times the parameters.

Training data improvements include tripling the number of RL environments, adding instruction-following overlays, and implementing a flywheel system powered by previous SWE-2 checkpoints that iteratively strengthens verification capabilities.

SWE-2 is immediately available through Devin Desktop and CLI interfaces, with planned rollouts for Devin Web and Fusion platforms. The release represents a crucial development in making advanced AI coding assistance more accessible through improved cost efficiency while maintaining competitive performance levels.

This advancement positions Cognition strategically in the rapidly evolving AI coding assistant market, where cost-effectiveness is becoming equally important as raw capability for widespread enterprise adoption. The focus on optimizing the entire cost-performance frontier rather than just peak performance suggests a maturing understanding of real-world deployment requirements in AI-assisted software development.