TECH Signal 133
Cognition’s SWE-2 reportedly tops Terminal-Bench 2.1 with 92.8 score at lower cost than Claude Fable 5.1
↗
Cognition’s SWE-2, a 2.8T-parameter MoE model, achieves a self-reported 92.8 on Terminal-Bench 2.1 while claiming 64% lower cost than Claude Fable 5.1 for comparable FrontierCode performance.
SWE-2 demonstrates that multi-trillion-parameter models can deliver competitive agentic coding performance at significantly reduced operational cost. However, its proprietary weights and lack of independent benchmark validation limit immediate adoption for engineers outside Cognition’s ecosystem. The gap in long-horizon tasks (Terminal-Bench 4.0) highlights where further development is needed.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
SWE-2 uses a 2.8T-parameter MoE architecture with 104B active parameters per token, post-trained from Kimi K3 with reinforcement learning at scale.
Self-reported benchmarks show parity with Claude Fable 5.1 on FrontierCode 1.1 (50.0 vs 50.9) at a claimed 64% lower cost, but lag significantly on Terminal-Bench 4.0 (27.3 vs 55.8).
No local inference or per-token API is available; SWE-2 is accessible only through Cognition’s Devin Desktop, CLI, and web platforms.
THE CLUSTER