Cognition's SWE-2 achieves 92.8 on Terminal-Bench 2.1

Cognition’s SWE-2, a 2.8-trillion-parameter MoE model with 104 B active parameters per token, achieves 92.8 on Terminal-Bench 2.1 and 50.0 on FrontierCode, matching Claude Fable 5.1’s score with 64% lower cost. The model uses MoE inference on NVFP4/FP8 kernels, adds 5–6 benchmark points over its Kimi K3 base, and improves token-per-second throughput by 10–20% with a prefill delayer.