The KodaCode team conducted a comparative analysis of agentic shells, which serve as interfaces for LLMs to interact with files and tests. The study utilized the DeepSeek V4 Flash model across various shells, including Koda, OpenCode, KiloCode, Claude Code, and Ouroboros, to evaluate how different implementations of the agentic cycle influence performance.
The evaluation was based on the 'Koda Index,' a metric comprising 261 tasks across three categories: software design, implementation, and data analysis. The benchmark includes 125 tasks from five sources, requiring agents to operate in isolated sandboxes without network access. Tasks are considered successful only if they pass automated verification tests.
Results showed that the Koda shell outperformed others, with a 2.7 percentage point lead over the nearest competitor at the Max reasoning level. The team also updated its proprietary models: Koda Base now uses Qwen 3.8 Flash 125B, while Koda Pro is built on GLM 5.3. Koda Pro achieved the highest score in the benchmark at 54.3%.
The Koda Base model, which remains free with daily limits, scored 49.5% in the benchmark. The average estimated cost per task for Koda Base is 11.9 ₽. The previous version of Koda Base was based on the Qwen 3.6-35b model, which the developers noted struggled with long agentic chains and tool usage.
