JetBrains AI Shifts to Agent-First Model

This new architecture provides developers with sensible defaults immediately applicable to most coding workflows, eliminating the need to configure agents like Junie, Codex, or Claude from the start.

During testing, Codex powered by GPT 5.6—an update to the previous 5.4 Mini—demonstrated the optimal balance of correctness, speed, and cost efficiency, leading the team to emphasize that it remains superior to Junie despite the release of newer models.

Despite OpenAI's release of GPT 5.6, the JetBrains team confirmed that Codex continues to outperform Junie in their evaluations, committing to ongoing assessment of new models to maintain optimal performance.

The JetBrains team evaluated multiple AI agents by testing them against a series of real-world tasks in programming languages such as Java, C#, and Python.

These tasks included bug fixes, feature development, and enhancements to existing codebases, with success measured by the agent's ability to solve the task, its operational cost, and its latency or speed.

The rigorous evaluation process ultimately narrowed down the contenders to two finalists: Codex with GPT 5.4 Mini and Junie with Gemini 3 Flash.

Codex demonstrated a slight advantage in its overall solve rate across the various tested languages, while Junie showed specific strengths in certain areas, proving to be faster and more cost-effective for particular tasks.

A-B testing revealed that Junie was more efficient for Java-based tasks, whereas Codex performed more effectively for projects in Python and C#.

Codex was ultimately selected as the default agent due to its consistent generalized performance across all tested languages, while users retain the freedom to switch to other agents if a different model better suits their specific project needs.