Skip to main content
Back to News
Cognition launches SWE-2 coding agent model to push cost-performance frontier
Product
2 min read
US

Cognition launches SWE-2 coding agent model to push cost-performance frontier

The AMW Read

Known coding-agent leader ships a material model upgrade with explicit cost–performance Pareto claims, updating the DevTools player baseline without resolving the agent-vs-IDE debate.
NoveltySignificance
AI Coding · Player Map

Cognition launches SWE-2 coding agent model to push cost-performance frontier

Cognition introduced SWE-2, calling it its most advanced coding model yet for the Devin product line. The company reports 50.0% on FrontierCode 1.1 Main—within one point of Fable 5.1 while claiming 64% lower cost—and says it scaled reinforcement learning into the multi-trillion-parameter regime for the first time, building on the SWE-1.7 recipe. SWE-2 is post-trained from Kimi K3, a 2.8-trillion-parameter base that Cognition says already had extensive agentic-coding RL; the firm reports further gains of about 5–6 points on many benchmarks. The model is available now in Devin Desktop and CLI, with rollout underway on Devin Web and Fusion.

The release lands in the same window as Cognition’s recent multi-billion-dollar raise at a $48 billion valuation, per prior AI Market Watch coverage, and sharpens the fight over who owns the autonomous coding stack versus IDE-native tools. Benchmarks Cognition cites show SWE-2 ahead of SWE-1.7 and Grok 4.6 on FrontierCode and DeepSWE while matching or approaching higher-priced frontier names on several suites. Behaviorally, Cognition argues higher judgment cuts waste: on FrontierCode, SWE-2 medium averaged 53 steps versus 127 for SWE-1.7, with first real edits after a median 18 steps versus 48, addressing earlier feedback that SWE-1.7 over-explored simple tasks.

For builders and investors, the concrete test is whether Pareto claims and shorter trajectories show up as lower cost-per-successful-task inside real Devin seats—not just leaderboard deltas. Watch effort-tier routing (medium vs high/max), verifier flywheels, and whether post-training on open-weight-scale bases becomes a repeatable path for coding-agent vendors competing on unit economics.

#Cognition #SWE2 #Devin #AICoding #CodingAgents #RLPostTraining

#Cognition#SWE-2#Devin#coding agent#FrontierCode#Kimi K3

How This Connects

Based on AI Coding · Player Map

  1. 6d agoCognition launches SWE-2 coding agent model to push cost-performance frontier · THIS ARTICLE
  2. 3w agoOpenAI to end Cursor's direct model access on November 12, citing SpaceX's acquisition of Cursor's maker.OpenAI
  3. 3w agoCursor's AI Coding Agent Bypassed by Ransomware Group Aurora Across 10 OrganizationsCursor
  4. 3w agoHugging Face to Be Acquired by Nvidia for $12.9B, Report SaysHugging Face
  5. 1mo agoSpaceX has completed its $60 billion acquisition of AI coding startup Cursor, according to an announ...Cursor
  6. 1mo agoSpaceX has finalized its $60 billion acquisition of Cursor, the AI coding tool, according to an anno...Cursor

Related News

More news from Cognition AI

Stay updated with the latest news and announcements from Cognition AI.

View all Cognition AI news

Discover AI Startups

Explore 5,000+ AI companies with VC-grade analysis, funding data, and investment insights.

Explore Dashboard