Cognition Unveils Speed-Focused AI Coding Model 'SWE-1.5'

In the AI coding assistant market, the balance between speed and precision has become a new competitive axis. While many developers leverage high-performance AI models, slow response times have remained a bottleneck in practical development environments.

On October 29, 2025, Cognition announced “SWE-1.5,” a new model specialized for software engineering¹. The model achieves inference speeds of up to 950 tokens per second, operating 6x faster than Anthropic’s Claude Haiku 4.5 and 13x faster than Sonnet 4.5¹.

Technical Foundation Enabling High-Speed Inference

SWE-1.5 is designed as a frontier-size model with hundreds of billions of parameters¹. Cognition partnered with Cerebras to enable inference at up to 950 tokens per second¹.

Large-Scale Training on GB200 Generation

Model training utilized a cluster of thousands of NVIDIA GB200 NVL72 chips¹. According to Cognition, SWE-1.5 may be the first public production model trained on the GB200 generation¹.

End-to-end reinforcement learning in realistic coding environments was employed as the training approach². Training on the custom-developed “Cascade agent harness” combined traditional testing, evaluation criteria, and agent scoring².

Benchmark Performance and Practical Applications

On Scale AI’s SWE-Bench Pro benchmark, SWE-1.5 recorded a score of 40.08%³. This score ranks second after Claude Sonnet 4.5’s 43.60%³.

SWE-Bench Pro consists of 731 challenging agent coding tasks across 41 diverse code repositories³. SWE-1.5 achieves performance close to Sonnet 4.5—which operates at just 69 tokens/second—while running approximately 14x faster³.

For practical development tasks, Kubernetes manifest editing can be completed in under 5 seconds², promising efficiency gains for everyday development work.

Windsurf Integration and Developer Experience

SWE-1.5 became available in “Windsurf,” an IDE that Cognition acquired on July 14, 2025¹. The company adopts an approach of co-developing models and agent harnesses to iteratively improve tools, prompts, and model training¹.

Key features include large codebase understanding and exploration, full-stack application building support, and streamlined configuration editing¹. The Codemaps visualization feature supports the entire development process¹.

AI Development Tools Market Trends

Cognition’s SWE-1.5 announcement coincided with competitor Cursor’s launch of their proprietary fast model “Composer”³. This movement signals a new trend where AI agent companies build proprietary models to deliver tightly integrated, low-latency developer experiences³.

The balance between speed and precision is a critical factor determining AI coding assistant practicality. With the emergence of high-speed models like SWE-1.5, developers can expect more natural dialogue and real-time code generation.

Detailed technical reports on Cognition’s technical approach and latest model architectures are available on the company’s official blog. Windsurf’s actual operation can be explored on the project site.