Cognition released SWE-1.6 on April 7, making it generally available in Windsurf with a free tier for the next three months. The model improves SWE-Bench Pro scores by over 10% compared to SWE-1.5 while running at up to 950 tokens per second on Cerebras hardware.
For the broader landscape, see our complete guide to AI coding tools in 2026.
What Happened
SWE-1.6 is Cognition's latest coding agent model, post-trained from scratch to optimize both intelligence and user experience. The model reduces unnecessary terminal reliance, uses parallel tool calls more often, and loops less frequently during complex tasks. Cognition calls this focus on "model UX" a new training objective alongside raw benchmark performance.
The release follows an early preview that showed the 10%+ SWE-Bench Pro improvement. The final model maintains that benchmark performance while adding behavioral refinements that make it smoother to use in continuous agent workflows.
Why It Matters
A free, fast coding agent lowers the barrier for creators who use AI to build tools, automate workflows, or develop plugins. The 950 tokens/second speed through Cerebras means near-instant responses, making it practical for interactive development sessions. The explicit focus on model UX addresses a common complaint: coding agents that technically solve problems but require constant babysitting.
Key Details
- SWE-Bench Pro: 10%+ improvement over SWE-1.5
- Speed: 200 tok/s (Fireworks, free) or 950 tok/s (Cerebras, paid)
- Free tier: 3 months, available now in Windsurf
- UX improvements: Fewer loops, more parallel tool use, less terminal reliance
What to Do Next
SWE-1.6 is available now in Windsurf. For a broader look at the coding AI landscape, see our comparison of AI coding tools in 2026.
