Cognition's New AI Coder Does the Same Job for 64% Less

Cognition, the maker of the AI coding assistant Devin, has launched a new model called SWE-2. It's designed to do the work of a junior-to-mid-level programmer: reading existing code, figuring out what's broken, writing fixes and running tests. On a widely used test of real coding tasks called FrontierCode, SWE-2 scored 50% — basically tied with a leading rival, Fable 5.1, which scored 50.9% but costs 64% more. Think of it as a contractor who delivers nearly the same quality of work for a third of the invoice.

The bigger story is efficiency. The previous version of Cognition's model took an average of 127 steps to finish a task; SWE-2's middle setting does it in 53. In plain terms, it wastes less time reading irrelevant files and re-checking its own work, and it starts writing actual code after about 18 steps instead of 48. The company says that adds up to 81% lower cost per task. Fewer steps also means less waiting — a developer sits idle far less while the AI churns.

How did they pull this off? SWE-2 is built on top of Kimi K3, an enormous model (2.8 trillion internal settings, sometimes called parameters). Cognition then used reinforcement learning — a training method where the AI is rewarded for good outcomes and penalized for bad ones — at a scale the company says is a first. Crucially, they also penalized the model for spending money, so it learned to be thrifty on its own.

The catch: SWE-2 is close to the best, but not the best. A rival model, GPT-6 Astra, scores 53.3% on the same test, though it costs roughly four times as much. Benchmark scores also don't capture messy real-world codebases. Still, for anyone paying for AI coding help, this is a clear signal: the same work is getting dramatically cheaper, fast.