Cognition SWE-2: A Devin Coding Model Built on Kimi K3

Cognition reports 50.0% for SWE-2 versus 50.9% for Claude Fable 5.1 on Cognition's FrontierCode 1.1 Main test, but 27.3% versus 55.8% on Terminal-Bench 4.

A coding model built on someone else's open weights

SWE-2 is Cognition's newest in-house coding model, following SWE-1.7 two months earlier. Cognition post-trained it from Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight model that had already been through extensive reinforcement learning for coding agents. The change Cognition highlights is a reinforcement learning method that trains every reasoning-effort level in a single run, so the low and high settings improve together. [1]

It shipped the same day in Devin's desktop app and command-line tool, with Devin's web app and its Fusion product to follow. It came two days after Cognition raised over $2 billion. [1][3]

Near the frontier on one test, far from it on another

On FrontierCode 1.1 Main, a benchmark Cognition publishes, SWE-2 scored 50.0 percent against 50.9 percent for Claude Fable 5.1 and 53.3 percent for GPT-6 Astra; Cognition says it gets there at 64 percent lower cost than Fable 5.1. On Terminal-Bench 4 it scored 27.3 percent, against 55.8 percent for Fable 5.1 and 57.9 percent for Astra. Cognition also says SWE-2 at medium effort beats SWE-1.7 while taking 58 percent fewer turns and costing 81 percent less. [1]

Atlas interpretation: Every number here is Cognition's, and the flattering one comes from Cognition's own benchmark. The Terminal-Bench gap is the more honest measure of distance from the frontier. The pitch is not that SWE-2 matches the best models everywhere but that it is good enough on typical coding work at a much lower price, which is the same bet Cursor made with its own models before SpaceX bought it. [1][4]

Sources

  1. Introducing SWE-2: Pushing the Pareto Frontier

    Cognition · Sep 10, 2026

  2. Cognition Releases SWE-2: A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost

    MarkTechPost · Sep 12, 2026

  3. Do it all with Devin: Announcing our Series E

    Cognition · Sep 8, 2026

  4. Cognition hits $48B valuation, signaling investors believe AI coding is far from a winner-take-all market

    TechCrunch · Sep 8, 2026