Cognition’s SWE-2 Targets the Cost-Performance Frontier for Coding Agents

Introduction

Cognition has released SWE-2, the latest model in its coding-agent lineup. The company is presenting it less as a simple attempt to set a new absolute benchmark record and more as an effort to improve the cost-performance frontier available to developers. Cognition reports a 50.0% result on FrontierCode 1.1 Main, roughly one percentage point behind Fable 5.1 while costing 64% less. SWE-2 is available in Devin Desktop and CLI, with rollout planned for Devin Web and Fusion.

Key points

• A large base model with additional post-training. SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter model that had already received reinforcement learning for agentic coding. Cognition says its own post-training adds roughly 5–6 points on many benchmarks compared with the base model.
• One run for several effort levels. The central technical change is a reward objective that combines task success with a penalty for rollout cost and time. The penalty is adjusted for each effort level according to the local slope of the base model’s cost-performance curve.
• More focused exploration. On FrontierCode 1.1 Main, Cognition reports an average of 53 steps for SWE-2 medium, compared with 127 for SWE-1.7. The median point at which the model makes its first substantive code edit reportedly falls from 48 steps to 18.
• Strong but uneven benchmark results. SWE-2 scores 50.0% on FrontierCode 1.1 Main, 73.0% on DeepSWE 1.1, and 92.8% on Terminal-Bench 2.1. Its 27.3% result on Terminal-Bench 4 is much lower than some competing systems, showing that model rankings depend heavily on task composition.

Why it matters

The notable aspect of SWE-2 is that Cognition treats intelligence and efficiency as linked optimization targets. A model that can identify the relevant parts of a repository sooner may solve straightforward tasks with fewer tool calls, while higher effort levels can still spend more time planning and verifying difficult changes. The company also describes expanded reinforcement-learning environments, instruction-following overlays, and a feedback loop in which earlier SWE-2 checkpoints help strengthen verifiers. Quantization and scheduling improvements are intended to reduce memory pressure and limit the mismatch between training and serving.

Cognition highlights three behavioral changes: broader end-to-end test coverage, more resourceful alternatives when an obvious route is unavailable, and stronger verification when the model is challenged. These observations come from the company’s own testing, so they should be treated as product claims rather than independent evidence. The spread between SWE-2’s benchmark results also argues against choosing a coding model from a single leaderboard number.

The broader implication is that coding-agent evaluation is moving beyond peak solve rate. As reasoning becomes more expensive, developers will care about how many tasks a model completes per dollar, how many steps it takes, and how quickly it starts useful implementation. SWE-2’s approach is an attempt to make those tradeoffs part of training itself; its long-term value will depend on whether the reported frontier improvements hold across independent workloads.

Source: Hacker News