Cognition releases SWE-1.6 with model UX improvements
Cognition Labs has released SWE-1.6, a software engineering model for Windsurf that prioritizes smoother agent behavior — including parallel tool calls, reduced looping, and less terminal reliance — alongside speeds up to 950 tok/s and a free tier for three months.
Score breakdown
Developers using Windsurf can now run SWE-1.6 for free and expect fewer interruptions from looping or terminal-heavy behavior, meaning the agent requires less manual intervention and completes tasks in fewer turns.
- 01SWE-1.6 is now generally available in Windsurf, free for all users for the next 3 months.
- 02The earlier SWE-1.6 Preview improved SWE-Bench Pro scores by more than 10% over SWE-1.5.
- 03SWE-1.6 was post-trained from scratch to optimize for both intelligence and model UX.
Cognition Labs has released SWE-1.6, a software engineering model post-trained from scratch to jointly optimize for raw intelligence and what the team calls "model UX." The model is now generally available in Windsurf. Unlike the earlier SWE-1.6 Preview — which was post-trained on the same pre-trained model as SWE-1.5 and improved SWE-Bench Pro scores by more than 10% — SWE-1.6 was rebuilt to address a set of behavioral friction points observed in the Preview checkpoint. These included overthinking on simple tasks, calling tools sequentially rather than in parallel, preferring shell commands over native tools, and getting caught in looping reasoning cycles.
Ablation results showed that task solve rate remained stable while assistant turn count stayed flat, indicating the model became more efficient without sacrificing capability.
The primary training intervention was the introduction of a length penalty, which discourages unnecessarily long trajectories. Ablation results showed that task solve rate remained stable while assistant turn count stayed flat, indicating the model became more efficient without sacrificing capability. In practice, SWE-1.6 now uses parallel tool calls more frequently, loops far less, and relies more on its own tools than the terminal — reducing the need for users to manually accept commands and allowing the agent to work more continuously.
SWE-1.6 is free for all Windsurf users for the next three months, delivered via a partnership with Fireworks at 200 tok/s. Paying users can access a faster version through a Cerebras partnership at up to 950 tok/s. The article was authored by Rohan Choudhury, Carlo Baronio, Ben Pan, Sam Lee, Eric Lu, Steven Cao, Joe Li, Andrew Wang, Adam Zweiger, Ray Wang, Gary Chang, and Silas Alberti.
Key facts
- 01SWE-1.6 is now generally available in Windsurf, free for all users for the next 3 months.
- 02The earlier SWE-1.6 Preview improved SWE-Bench Pro scores by more than 10% over SWE-1.5.
- 03SWE-1.6 was post-trained from scratch to optimize for both intelligence and model UX.
- 04Key behavioral issues addressed: overthinking, sequential tool calls, terminal over-reliance, and looping.
- 05A length penalty was introduced during training to discourage unnecessarily long trajectories.
- 06Fireworks partnership delivers the free tier at 200 tok/s; Cerebras partnership delivers a paid tier at up to 950 tok/s.
- 07Ablations showed task solve rate stayed stable while assistant turn count remained flat.