Cognition SWE-1.7 runs at 1,000 tok/s
Cognition's SWE-1.7 scores 42.3% on FrontierCode 1.1 and runs at 1,000 tok/s via Cerebras at $1.97 per task, built on Kimi K2.7-Code.
Cognition released SWE-1.7 on July 8, trained from Moonshot AI’s Kimi K2.7-Code with additional reinforcement learning post-training inside the Devin task harness. The model scores 42.3% on FrontierCode 1.1 and 81.5% on Terminal-Bench 2.1, a 12.2-point gain over the base Kimi K2.7-Code at 30.1% on FrontierCode.
The method matters here. Cognition applied RL to a model that had already undergone heavy RL training, which the team calls a challenge to the idea of a post-training ceiling. The gains kept coming rather than flattening out. A third benchmark, SWE-Bench Multilingual, puts SWE-1.7 at 77.8%.
For context, FrontierCode 1.1 places SWE-1.7 at 42.3%, just below GPT-5.5 at 43.0% and Claude Opus 4.8 at 46.5%. Where SWE-1.7 separates itself is cost and speed: it runs at 1,000 tokens per second via Cerebras and comes in at roughly $1.97 per task on the FrontierCode Main benchmark set, compared to substantially higher rates for frontier models.
SWE-1.7 is available now through Devin’s web interface, desktop app, and CLI. For background on the base model, see the Kimi K2.7-Code release. For the frontier end of the comparison, Claude Sonnet 5 is the current reference point. The model routing guide covers when the price-performance trade-off favors a specialist model.