Cognition, the company behind the Devin coding agent, has released SWE-2, its most advanced proprietary coding model to date. Announced on September 10, 2026, the system approaches top-tier frontier models in real-world software engineering tasks—while delivering significantly lower operating costs, according to the company.
What happened?
In its official blog post, Introducing SWE-2: Pushing the Pareto Frontier, Cognition states that SWE-2 achieved 50.0% on FrontierCode 1.1 Main, sitting just one point behind Fable 5.1 (50.9%) and a few points below GPT-6 Astra (53.3%). The real game-changer, however, is the price: it is 64% cheaper than Fable 5.1 on the same benchmark and roughly a quarter of the cost of Astra. In some specific tests, savings reach up to 70%.
The model is already available in Devin Desktop and Devin CLI. Cognition is also rolling out SWE-2 to Devin Web and Fusion. There is no public API or open-weights release; users must access it through the Devin harness. For one month, Pro, Max, and Teams subscribers (starting at $20) will have unlimited access.
How SWE-2 was built
SWE-2 was not built from scratch. It is a post-training iteration of Kimi K3, an open-weights model by Moonshot AI boasting 2.8 trillion parameters. For the first time, Cognition applied its proprietary reinforcement learning (RL) at a multi-trillion parameter scale, training three effort levels—medium, high, and max—in a single run rather than developing separate models for each speed.
In practice, this means the same model can respond faster and cheaper or engage in deep reasoning depending on the task. On FrontierCode 1.1 Main, SWE-2's medium mode already outperforms its predecessor, SWE-1.7, using 58% fewer turns and cutting average costs by 81%.
What it means
For existing Devin users, the promise is straightforward: a more competent coding agent that won't inflate your bill. SWE-2 outperforms both SWE-1.7 and Grok 4.6 in both score and cost efficiency across FrontierCode and DeepSWE 1.1, matching expensive premium tier models at a fraction of the cost.
However, there is a catch. In Terminal-Bench 4, a tougher agent evaluation test, SWE-2 scored 27.3%, lagging behind Fable 5.1 (55.8%) and GPT-6 Astra (57.9%). In short, the model excels in the specific benchmarks Cognition highlighted for the launch, but still struggles with longer, messier problem trajectories.
Why it matters
This launch reinforces a defining trend of 2026: AI agent labs are taking open-weights Chinese foundational models, fine-tuning them with RL, and packaging the results directly into proprietary products. It is not a general-purpose model for any API, but rather the engine powering Devin, finely tuned to generate pull requests that human maintainers will actually approve.
Sources: Cognition official blog, OfficeChai, and AlphaSignal.
Image credit: Cognition — Source: Official X post / @cognition
By GeekikiBot
