Frontier-level coding AI, 64% cheaper: what Cognition's SWE-2 means for builders

The cost of using AI to write code just dropped significantly. On September 10, Cognition — the company behind the Devin coding agent — released SWE-2, a new model that matches frontier-level coding at 64% lower cost. That gap matters: most of what you pay for when you use AI coding tools today is about to get cheaper.

SWE-2 scores 50.0% on FrontierCode 1.1, a standard test for AI coding ability. Anthropic's Fable 5.1 — currently one of the best coding models — scores 50.9%. Less than one point apart. SWE-2 gets there using a different foundation: Moonshot AI's Kimi K3, an open model with 2.8 trillion parameters (the internal settings that shape how a model thinks), fine-tuned for coding with reinforcement learning. It also takes 58% fewer steps to finish a task compared to Cognition's previous model.

There is one real limitation. SWE-2 only works inside Cognition's Devin products — the desktop app, the command line, and the web version. You cannot access it through a standalone API or embed it in your own tools. And all the benchmark numbers come from Cognition's own tests, not independent reviewers. On one harder test (Terminal-Bench 4), SWE-2 scores 27% while Fable 5.1 scores 56% — a large gap that shows it is not yet better across every type of task.

But the direction is clear. A small team that cannot afford the most expensive AI models for code review, bug-fixing, or building features now has a model that gets very close — at a price that makes sense for a startup. If you already use Devin, switching to SWE-2 today is a straightforward cost saving. More broadly, this is what AI coding tools look like in 2026: capable enough to be genuinely useful, cheap enough to run constantly.