50.0% versus 50.9% — a 0.9-point gap, at 64% lower cost. That is the headline from Cognition's September 10 release of SWE-2: on its own FrontierCode 1.1 Main benchmark, the model sits on nearly the same line as Anthropic's Fable 5.1 while costing a fraction of the price.
The re-training dividend of an open base
SWE-2 is not built on an in-house foundation model. Its base is Kimi K3, Moonshot AI's 2.8-trillion-parameter open-weight model — one that had already undergone extensive RL for agentic coding. According to Cognition's official post, its RL recipe still found substantial headroom on this already-trained model: 5–6 additional points on many benchmarks, shifting K3's entire cost–performance frontier. By the company's own account, this is also the first time the team has scaled RL post-training into the multi-trillion-parameter regime.
The methodological highlight is that a single RL run now covers every reasoning-effort level. The reward reads R = S − λe·C, where S is task success, C is rollout cost (a mix of inference cost in USD and time), and λe is tuned to the slope of the base model's Pareto frontier at each effort level. The practical consequence: the model cannot profit by "downshifting" — at high effort, pretending to be the medium tier saves cost but does not increase reward; only genuinely pushing the frontier up does. The surrounding engineering is equally dense: a length-weighted reward baseline keeps training KL stable; DSpark speculative decoding plus a SpecForge-trained draft model lengthens accepted sequences by 15% while tracking the policy online; NVFP4/FP8 kernels with quantization-aware training rein in memory. The behavioral shift is visible: SWE-2 medium makes its first real edit after a median of 18 steps versus 48 for SWE-1.7, takes 58% fewer turns, and costs 81% less on average — over-exploration is replaced by knowing which part of the codebase actually matters.
The scorecard, and the cold water
The official table lists four benchmarks. FrontierCode 1.1 Main: 50.0% (Kimi K3 44.2, Grok 4.6 48.0, GPT-5.6 Sol 47.5, Fable 5.1 50.9, GPT-6 Astra 53.3, SWE-1.7 42.0). DeepSWE 1.1: 73.0%, second only to GPT-6 Astra's 74.1%. Terminal-Bench 2.1: 92.8% — the only row where SWE-2 beats every model listed. The cold water sits in the fourth row: Terminal-Bench 4 comes in at 27.3%, against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra. Independent analysis states plainly that "on par with the frontier" holds on three benchmarks and not the fourth; that long-horizon terminal gap is the footnote most worth remembering inside the "64% cheaper" story.
So what
SWE-2 is live in Devin Desktop and CLI, and rolling out to Devin Web and Fusion. For the industry, the significance is less any single score than the path it re-validates: the value of open weights is not just free access, but that a third party can post-train them into a commercial model that closes on the frontier with a published methodology. When a 2.8T open base meets a team that knows how to train the whole Pareto frontier, how much pricing power do closed labs have left?
References: Cognition's official post; cellcog.ai benchmark breakdown; AI/TLDR digest.