3code benches 75 lower token use vs. OpenCode on 10-task SWE-Bench subset (capocasa.dev)

In a significant development for the AI/ML community, the recently validated coding agent 3code has demonstrated a remarkable 75% reduction in token usage compared to its counterpart, OpenCode, during a benchmark of ten tasks from the SWE-Bench subset. Both systems utilized the same underlying model, GLM 5.2, illustrating the efficiency of 3code in executing programming tasks. The initial results not only highlight a more effective use of resources but also show that 3code succeeded in tackling a task that OpenCode could not, suggesting superior performance in practical applications. The implications of this validation are substantial, particularly as 3code's design included fundamental optimizations, such as verbal compacting of prompts and modular system architecture. These adjustments are described as "low hanging fruit" that could be further enhanced through additional features like 'cybernetic mode,' which is not yet autoloaded in 3code. As it stands, this early achievement points to the potential for greater efficiency and effectiveness in coding agents, which could revolutionize how programming tasks are automated and performed in the industry.