Cognition releases SWE-1.7, claiming near-frontier coding at about $1.97 per task

Cognition released SWE-1.7 on July 8, 2026, an agentic coding model the company says approaches frontier performance while costing about $1.97 per task through Cerebras inference.

The release lands in an intensifying price war among coding agents, where labs now compete on cost per task as much as raw capability. Cognition trained the model from a Kimi K2.7 base, added reinforcement-learning post-training, and runs it through its coding agent, Devin, at roughly 1,000 tokens per second on Cerebras hardware.

On Cognition’s own FrontierCode 1.1 benchmark, SWE-1.7 scores 42.3%, just behind GPT-5.5’s 43.0% and further behind Claude Opus 4.8’s 46.5%. It trails both leaders on Terminal-Bench 2.1, at 81.5% against 86.9% for Opus 4.8, and reaches 77.8% on SWE-Bench Multilingual, ahead of GPT-5.5 but well behind Opus.

The company credits the gains to several training changes, among them preserving policy entropy to stabilize learning, training across clusters on three continents with fault tolerance, and a “self-compaction” technique that extends how long the model can work on a task beyond its context window.

The benchmark figures are Cognition’s own and have not been independently verified; the model trails Opus 4.8 on every test the company reported, and the per-task cost comes from launch coverage rather than a published price list.