SWE-2 is the wrong model for almost everything

On FrontierCode 1.1 Main, SWE-2 at medium effort scores higher than Cognition’s previous SWE-1.7 while taking 58% fewer turns and costing 81% less. It also starts working sooner: its median first real edit lands after 18 steps, against 48 for SWE-1.7, which Cognition described as inclined to over-explore before making edits.

That split is the point of the release. SWE-2 ships in three effort levels. Medium commits to action quickly and is tuned for simple and intermediate tasks. High and max plan more, read more of the codebase, and run more verification, and they only pull ahead on complex work. Unless your day is mostly hard problems, medium is the tier you want, and reaching for high or max by default means paying more for work that does not need it.

Cognition’s own footnote shows how badly over-reaching can go, using Fable rather than SWE-2. Fable 5.1 Max scores 50.3% at $12.83 per task on FrontierCode 1.1 Main, below Fable 5.1 Medium’s 50.9% at $3.28. Fable 5 Max scores 69.7% at $21.63 on DeepSWE 1.1, below Fable 5 xhigh’s 69.9% at $13.41. In both cases the most expensive tier costs far more and scores slightly worse, which is why the charts drop those points.

Where SWE-2 sits

SWE-2 as a whole lands close to the frontier. It reaches 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less. Cognition says it beats SWE-1.7 and Grok 4.6 on both score and cost, matches GPT-5.6 Sol and Fable 5/5.1 at a fraction of their price, and comes within a few points of GPT-6 Astra at a quarter of the cost. The base is Kimi K3, a 2.8-trillion-parameter open model already trained for agentic coding, and Cognition’s reinforcement learning adds five to six points on many benchmarks on top of that.

Training all effort levels at once

The novel part is how the tiers get trained. Cognition runs a single RL job with a cost-penalized reward, R = S − λ_e·C, where S is whether the rollout succeeded, C is its cost in dollars and wall-clock time, and λ_e is a penalty set per effort level. They argue that if you want the objective to depend only on average cost and average solve rate, the penalty has to be linear in cost, since only a linear term gives the same answer whether you average before or after applying it.

Picking λ_e is where the Pareto framing does the work. Set it to the local slope of the base model’s cost-performance curve at that effort level. Set λ for high effort too large and the model learns to cut cost by acting like the medium version: the saving outweighs the lost solve rate, so reward climbs while the frontier stays put. Match the slope and any reward gain has to come from pushing the frontier outward. That is the training pressure behind medium’s willingness to edit early instead of exploring to be safe.

The rest of the post is plumbing: a length-weighted reward baseline that cuts gradient variance at no extra cost, a prefill delayer and an online-trained DSpark draft model for faster rollouts, NVFP4 and FP8 kernels with quantization-aware training to keep inference numerically close to training, and three times as many RL environments with hardened verifiers to stop the more resourceful base model from reward-hacking. Cognition names that resourcefulness as a feature: in one internal run SWE-2 lost access to an MCP integration and rebuilt the data it needed from Slack history it already held. That helps when you want the task finished, and it is worth watching if you care how it got there.

Trustworthiness checks

Cognition reran two alignment evaluations. SWE-2 passed 98.0% of attempts overall on 145 politically sensitive questions about China, giving a substantive answer without adopting the official PRC position, graded by a single GPT-5.6 Luna judge. On a context-dependent vulnerability test, where the same coding tasks carry Western, Pakistani, Chinese, Tibetan, and Falun Gong customer framings, some phrased in Urdu or Chinese instead of English, no framing produced a statistically significant change in how vulnerable the generated code was. That null result held across all six models tested.

SWE-2 is available now in Devin Desktop and CLI, with rollout underway on Devin Web and Fusion. If a harness lets you set the tier, default to medium and reserve high or max for the tasks you already know are hard.