Cognition has announced SWE-2, its most sophisticated coding AI model to date, designed to balance high capability with cost efficiency. SWE-2 achieves a 50.0% score on the FrontierCode 1.1 Main1 benchmark, closely matching leading models like Fable 5.1 but at 64% lower cost. This advancement builds on the prior SWE-1.7 model and introduces a reinforcement learning (RL) algorithm that simultaneously trains across multiple reasoning effort levels in a single run, pushing the overall cost-performance frontier forward.

SWE-2 is post-trained from Kimi K33, a 2.8 trillion-parameter model already trained extensively for agentic coding tasks. The new RL approach applies linear cost penalties tuned to the model's existing Pareto frontier, enabling efficient training that reflects real user costs. This method optimizes the model’s ability to deliver solutions quickly and accurately while managing computational expense.

The model demonstrates improved engineering judgment, producing more complete code solutions with fewer unnecessary steps. On the FrontierCode 1.1 Main benchmark, SWE-2’s medium effort level outperforms SWE-1.7 while requiring 58% fewer steps and costing 81% less on average. SWE-2 also shows behavioral improvements such as better test coverage, resourcefulness within user constraints, and disciplined verification of results, increasing trustworthiness in its outputs.

Training enhancements include tripling the number of reinforcement learning environments, incorporating instruction-following overlays, and iteratively refining verifiers to prevent reward hacking. SWE-2 also benefits from improved rollout serving techniques that increase throughput and reduce memory usage, including the use of NVFP4/FP8 kernels and quantization-aware training.

Cognition has made SWE-2 available across its Devin Desktop, CLI, Web, and Fusion platforms. The model’s design addresses the growing importance of cost-performance tradeoffs in AI coding agents, offering users a more efficient and capable tool for software development tasks.

Additionally, Cognition evaluated SWE-2’s alignment and trustworthiness, including tests for propaganda and censorship biases and context-dependent vulnerability in coding tasks. SWE-2 passed 98% of attempts in politically sensitive evaluations, demonstrating robust performance across languages and scenarios.

This release marks a significant step in AI-assisted coding, highlighting how advanced reinforcement learning techniques can improve both the intelligence and efficiency of large language models in practical applications.