Cognition has launched a new AI coding model, SWE-1.7, now live in Devin. The model is built on Moonshot AI's Kimi K2.7 Code. In multiple agentic coding benchmarks, it approaches the performance of GPT-5.5 and Claude Opus 4.8 but at a fraction of the generation cost. By plugging into Cerebras inference chips, SWE-1.7 cranks out 1,000 tokens per second.
To handle marathon tasks, the model learns to automatically summarize its state at critical context windows and resume from that summary. This allows a single session to run for up to 6 hours. The training team introduced an alternating length penalty that discourages verbosity on simple problems while saving long-chain reasoning for truly hard ones. To escape the limits of any single cluster, they built a distributed training architecture spanning three continents — syncing only weight deltas, so global parameter updates take just minutes.