Cognition’s SWE-1.7 rivals GPT-5.5 at 1,000 tokens per second
- Source
- Cognition
- Time
- 2:17 AM
- Weight
- 94/100
Cognition has announced the release of SWE-1.7, a specialized model for agentic software engineering that achieves frontier-level performance at a significantly reduced cost. Trained on a Kimi K2.7 base, the new model rivals the capabilities of GPT-5.5 on key coding benchmarks, including FrontierCode 1.1 and Terminal-Bench 2.1.
It is currently available through the Devin platform and is powered by Cerebras hardware to deliver inference speeds of 1,000 tokens per second. The model’s development involved several technical innovations, such as a multi-cluster training infrastructure spanning three continents and a self-compaction technique that allows the agent to summarize its working state to handle long-horizon tasks.