Top AI Product

The Devin team shipped SWE-1.7 on July 8, their strongest coding model yet. It’s not a chatbot and not an app — it’s the agentic model that drives Devin’s autonomous software engineering: give it a task, it reads the repo, edits files, runs the terminal, and closes the loop on long-horizon work.

The number that matters

42.3% on FrontierCode 1.1, right behind GPT-5.5 (43.0%) and Opus 4.8 (46.5%), and well ahead of GLM-5.2 and Composer 2.5. Around 81.5% on Terminal-Bench. The catch that makes it interesting: roughly $1.97 per task, a sliver of what the frontier models cost.

How they got there is the real story. Cognition took the open-source Kimi K2.7 base from Moonshot and ran RL post-training on top — inside the actual Devin harness, across four data centers on three continents. RL on top of someone else’s RL, standing on an open model instead of training from scratch.

How you use it

There’s no direct API. SWE-1.7 lives inside Devin — Web, Desktop, and CLI — served on Cerebras at 1000 tokens/second. That speed is the point for an agent doing dozens of async edits.

You Might Also Like

• Kimi k2 6 Beats gpt 5 4 and Claude Opus 4 6 on swe Bench pro
• Gpt 5 6 sol on Cerebras 750 Tokens s Openais Frontier Model Gets 15x Faster in July
• Grok 4 5 Spacexai Solves swe Bench pro Tasks With 4 2x Fewer Tokens Than Opus 4 8
• Mercury 2 Just hit 1000 Tokens per Second and its not Even Using Transformers
• A Single api String Exposed Cursors Secret Composer 2 Runs on Moonshot ais Kimi k2 5