Cognition has released SWE-1.7, a software engineering model that achieves performance comparable to GPT-5.5 and Opus 4.8 while significantly reducing costs. The model, built on a Kimi K2.7 base, demonstrates that reinforcement learning can overcome previous 'post-training ceilings' through improved infrastructure and long-horizon task optimization. It is now available via Devin and Cerebras, showing strong results on benchmarks like SWE-Bench and Terminal-Bench.
Background
Cognition is developing AI agents for software development, with their flagship product being Devin. This release highlights the ongoing competition in specialized coding models and the push for more efficient, high-capability AI agents.
- Source
- hackernews
- Published
- Jul 9, 2026 at 12:19 AM
- Score
- 8.0 / 10