Cognition Unveils AI Models That Speed Up Code Search by 10x While Maintaining Accuracy

Cognition launches SWE-grep AI models that accelerate code search by 10x while maintaining accuracy, using reinforcement learning to execute parallel tool calls and reducing context retrieval time from 60% to under 5 seconds for developers.

• Cognition introduces SWE-grep and SWE-grep-mini, AI models that perform context retrieval for coding tasks an order of magnitude faster than frontier models while matching their accuracy
• The models use reinforcement learning to execute up to 8 parallel tool calls across 4 turns, reducing the typical 60% of time agents spend on context retrieval in their first turn
• Fast Context subagent now powers Windsurf's code search capabilities and aims to keep developers in flow state by responding within a 5-second window