Cognition ships SWE-2: near-frontier coding at a quarter the cost
Built on Moonshot's open Kimi K3, it matches Fable 5.1 for 64% less and is free in Devin for a month. Plus: Suno's licensed v6, Magic's cheap frontier run.
Copy markdown
The numbers: near-frontier, a quarter the cost
SWE-2 scores 92.8% on Terminal-Bench 2.1 and 73.0% on DeepSWE 1.1, and lands within a point of Anthropic's Fable 5.1 on FrontierCode (50.0 vs 50.9). Cognition says it runs at roughly a quarter of GPT-6 Astra's price and 64% cheaper than Fable at the same score.
Built on Moonshot's open Kimi K3
SWE-2 is post-trained from the open 2.8-trillion-parameter Kimi K3, which Cognition says is the first time RL has been scaled to the multi-trillion-parameter regime. The takeaway for builders: frontier coding is increasingly assembled on open weights you can also download.
58% fewer turns to the first edit
SWE-2 medium reaches its first code edit after a median of 18 agent steps versus 48 for SWE-1.7, using 58% fewer turns overall at 81% lower average cost. Tighter loops mean cheaper autonomous runs and less time watching Devin think.
Live in Devin now, free for a month
It's rolling out across Devin Desktop, CLI, Web, and Fusion in three effort tiers - medium, high, and max - and is free for all Pro, Max, and Teams subscribers for the next month. No standalone API or Windsurf pricing has been published yet.
Build this weekend: a budget fix-it agent
At a quarter of frontier cost and 18-step first edits, SWE-2 medium makes it cheap to point Devin at a backlog of failing tests, lint errors, or a big dependency bump and let it grind - the kind of long, boring job that was too expensive to automate at frontier rates.
Elsewhere: Suno's v6 goes fully licensed
Suno rebuilt its music models from scratch on licensed Warner, BMG, and Believe catalogs, shipping v6, v6-wild, and v6-mini (mini is free for everyone) and retiring the old models. Cleaner provenance for anyone generating soundtracks inside a product.
Elsewhere: Magic matched a frontier base for ~$500K
Magic says it reproduced DeepSeek V4 Pro base quality with about 50x less compute for roughly $500K - a sign that near-frontier pretraining is getting far cheaper, and that more capable open bases (like the one under SWE-2) will keep arriving.