Cognition's SWE-1.7 bets the ceiling is fake
Cognition dropped SWE-1.7 today, the coding model that runs inside Devin. The headline claim is not that it beats everyone, it doesn't. On FrontierCode 1.1 it hits 42.3% against GPT-5.5's 43.0% and Opus 4.8's 46.5%, on Terminal-Bench 2.1 it trails both, on SWE-Bench Multilingual it edges past GPT-5.5 but sits well under Opus. Frontier-adjacent, not frontier.
So why care. Two reasons, and neither is the score. First, Cognition says the whole point is that pushing reinforcement learning far past normal post-training keeps paying off, that the capability ceiling everyone assumed was there isn't. A smaller model kept climbing under more RL. If that holds, it changes the math on whether you need a giant base model at all. Second, and this is the one people will feel, SWE-1.7 runs at 1000 tokens per second on Cerebras inside Devin. That is roughly ten times the speed of the frontier models it competes with.
Put Grok 4.5 and SWE-1.7 side by side, both shipped today, and the pattern is loud. The coding-model fight has moved off the leaderboard. Grok 4.5 competes on price, SWE-1.7 on speed, and both are content to sit a few points below Opus because for an agent grinding through a task list, a model that is 90% as smart but 10x faster or 4x cheaper wins every time.
The assumption that you always reach for the single smartest model is quietly dying. Read the technical writeup at cognition.com/blog/swe-1-7.
← Back to all articles
So why care. Two reasons, and neither is the score. First, Cognition says the whole point is that pushing reinforcement learning far past normal post-training keeps paying off, that the capability ceiling everyone assumed was there isn't. A smaller model kept climbing under more RL. If that holds, it changes the math on whether you need a giant base model at all. Second, and this is the one people will feel, SWE-1.7 runs at 1000 tokens per second on Cerebras inside Devin. That is roughly ten times the speed of the frontier models it competes with.
Put Grok 4.5 and SWE-1.7 side by side, both shipped today, and the pattern is loud. The coding-model fight has moved off the leaderboard. Grok 4.5 competes on price, SWE-1.7 on speed, and both are content to sit a few points below Opus because for an agent grinding through a task list, a model that is 90% as smart but 10x faster or 4x cheaper wins every time.
The assumption that you always reach for the single smartest model is quietly dying. Read the technical writeup at cognition.com/blog/swe-1-7.
Comments