New SWE-1.5 Fast Agent Model Officially Introduced by Cognition: Now Available on Windsurf Code Editor

Today on the 30th of October 2025, one among the notable AI firms – Cognition officially introduced its new SWE-1.5 fast agent model. It has also been made available within the Windsurf code editor.

Check out more about it below.

Cognition’s New SWE-1.5 Fast Agent Model – Now Available on Windsurf

The newly-introduced SWE-1.5 fast agent model has been designed to offer both high performance and high speed in software engineering tasks, and thanks to its collaboration with Cerebras, this new model is capable of processing up to 950 tokens per second. In comparison to other models like Claude’s Sonnet 4.5 and the Haiku 4.5, SWE-1.5 is respectively 13x and 6x more faster.

Speaking more, Cognition’s SWE-1.5 is focused on offering end-to-end user experience, and it is revealed to have achieved near-SOTA coding performance at high speed. It should also be noted that the SWE-1.5 model is not just a standalone model, the model, inference process, and agent harness are all co-designed together to achieved both speed and intelligence, and work as a single system.

In terms of its training methodology, the training was carried forward using end-to-end reinforcement learning on Cognition’s custom Cascade agent harness, leveraged on a cluster of thousand GB200 NVL72 chips. Additionally, Cognition also manually generated a dataset of real-world tasks for better reflecting such scenarios. About its use cases, SWE-1.5 can be used to explore and understand large code bases, build end-to-end full-stack applications, and edit configurations.

Also, on the SWE-Bench Pro benchmark developed by Scale AI, Cognition’s SWE-1.5 model attained good performance scores of 40.08% (and it is positioned after Claude’s Sonnet 4.5 which achieved 43.60% performance score). However, the brand adds that many engineers use it as their daily driver for completing tasks.