Cognition releases SWE-1.7 coding model at $1.97 per task

Category: Agentic coding
Status: confirmed
Sources: Cognition blog, HN discussion
Summary: Cognition launched SWE-1.7 on 2026-07-08, its most capable coding model, reinforcement-learning trained on the open-weight Kimi K2.7 base. Cognition reports 42.3 percent on its FrontierCode 1.1 benchmark, which scores whether a maintainer would merge a produced pull request, against 43.0 percent for GPT-5.5 and 46.5 percent for Claude Opus 4.8. It states a cost of $1.97 per task on the FrontierCode Main set and runs the model at 1,000 tokens per second through Cerebras inside Devin (Web, Desktop, and CLI). Cognition says its RL training spanned four datacenters across three continents combining its own GPUs with inference-provider compute.
Why it matters: A near-frontier coding model at a fraction of the per-task cost of the largest models, trained on open weights, extends the pressure that cheaper models are putting on frontier-lab inference pricing.
Follow-up: Track independent reproduction of the FrontierCode figures, since the benchmark is Cognition's own, and availability outside Devin.