SWE-1.7 Reaches Near GPT-5.5 and Opus Intelligence — Frontier Coding at $1.97 Per Task

What happened

Cognition on July 8 released SWE-1.7 — its most capable coding model to date, trained with extensive reinforcement learning on a Kimi K2.7 base, exhibiting significantly different behavior from the base model.

Context and impact

SWE-1.7 targets long-horizon enterprise coding tasks lasting up to 6 hours. Running on Cerebras at 1,000 TPS enables dramatically lower cost vs. top frontier models. It scores 42.3% on FrontierCode 1.1 vs. GPT-5.5 (43.0%) — at >90% cost savings.

Details

• FrontierCode 1.1: 42.3% (GPT-5.5: 43.0%; Claude Opus 4.8: 46.5%)
• Terminal-Bench 2.1: 81.5%
• SWE-Bench Multilingual: 77.8%
• Cost: $1.97/task via Devin
• Speed: 1,000 tokens/s (Cerebras)
• Context: 256K tokens
• Techniques: multi-cluster training, self-compaction, online RL