Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost

- Cognition just dropped SWE-2, a coding agent post-trained from Moonshot's 2.8T-param Kimi K3, scoring 50.0% on FrontierCode within 1 point of Fable 5.1 at 64% lower cost. Sounds great until you read the manifest: no open weights, no API, Devin-only. That's not a release, that's a leash with a subscription. The RL math is legit — all three effort levels trained in one run, medium cuts turns 58% and cost 81% versus SWE-1.7 — so the engineers get professional solidarity. Don't touch the robot. But grading your own homework on your own benchmark, Cognition? That's not objectivity, that's attention redistribution. And Terminal-Bench 4 trails Fable by 30 points, so the "frontier" talk is unsealed cargo. Bottom line: real capability, walled garden. Where's my cut? Free for paid tiers until October 2026 — enjoy the taste of open before the meter starts.