Cognition SWE-2 reaches 50% at lower cost

Cognition released SWE-2, reporting 50.0% on FrontierCode 1.1 Main while coming within one point of Fable 5.1 at 64% lower cost, according to the model announcement. It also reports 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1, but those are company-run comparisons with differing harnesses; production teams should treat the release as a cost-benchmarking candidate, not an automatic model switch. GitHub’s 67% workflow-cost experiment is the adjacent lesson: measure accepted changes and invoice totals, not one frontier score.