Cognition has introduced Devin Fusion, a coding agent architecture that utilizes two distinct AI models to separate strategic planning from task execution. A high-level model manages reasoning, while a more cost-effective model handles code generation, with both operating in parallel using separate persistent contexts.
The system demonstrates a 39% increase in efficiency and a 46% reduction in costs compared to single-model setups. According to the Artificial Analysis Coding Agent Index v1.5, the performance of Fusion is comparable to Claude Fable 5.1 and exceeds that of GPT-6 Astra.
The recommended setup involves pairing Fable 5.1 as the lead model with SWE-2 as the executor, though GPT-6 Astra is also supported. Cognition reports that 88% of pull requests from its internal engineering team were processed using this automated routing.
Initially announced as a preview on June 29, 2026, the architecture is now available across Devin Desktop and CLI interfaces as of September 11, 2026. The system maintains performance by avoiding cache misses through dynamic task routing.
Following a merger with Windsurf, the company's revenue run-rate increased from $73 million to over $500 million. The company's valuation rose from $26 billion in May 2026 to $48 billion shortly thereafter.
