JetBrains Is Moving Beyond AI Chat, Here’s What Comes Next
In a Nutshell
JetBrains now defaults AI to agent mode with Codex (GPT 5.4 Mini) as the recommended agent. Codex was chosen after benchmarks showed it edges out Junie in solve rate, with acceptable cost and speed across Java, C#, and Python tasks. Users can still switch agents anytime, and JetBrains will keep re-evaluating models for the best balance of performance, cost, and latency.
These notes were generated by AI and may contain inaccuracies.
JetBrains moved to an agent-first setup and made Codex the recommended agent for JetBrains AI. Previously, AI users would start out in chat mode, then choose which coding agent they'd like to use and which model they want. This could be anything from Junie, Codex, Claude Agent, or another ACP-compatible agent. Now JetBrains AI starts off in agent mode, with sensible defaults that work well for developers in most use cases.
They are recommending Codex with GPT 5.4 Mini as it's shown to be a good balance between correctness, speed, and cost efficiency. Editor Mike from the future notes that since the recording, OpenAI has released GPT 5.6. While the initial tests were with 5.4, Codex still went out when compared to Junie. In the IDE, you'll see 5.6 powering Codex, but everything in this video is still relevant. Codex is recommended as a good default, though users can pick any agent they want.
JetBrains will continue to evaluate new models as they're released in order to make sure they're recommending the most effective and useful setup. Codex was selected after evaluating multiple agents against a series of real-world tasks in Java, C#, and Python. Tasks included bug fixes, feature development, and general enhancements across their test code base. The changes were then run against a test suite to check whether the agents actually did the work correctly.
The evaluations were broken down into three main questions:
- Can the agent actually solve the task?
- Is the cost reasonable?
- Is it fast enough?
The focus was on finding the best balance of solve rate, cost, and latency for real developer workflows, not just which model was the smartest or fastest. Models like GPT 5.5 or Fable may give better results but could be far more expensive.
The two finalists were Codex with GPT 5.4 Mini and Junie with Gemini 3 Flash. The benchmarks were really close. Codex had a slight edge in overall solve rate, while Junie was often faster and cheaper in certain areas.
An A-B test was run with real users, looking for signals like activation, churn, failure rate, and whether users switched away from the agent. Junie tended to be cheaper and more efficient in Java, with Codex being more efficient for Python and C#. The generalized nature of Codex helped tip the decision.
Codex is not the only agent users should use. Users can switch to another agent at any time. If in a Java-heavy project, switching to Junie might be preferred. If another agent already fits the workflow better, that one can be used as well. The goal is to make JetBrains AI more agent-first, with sensible defaults while still giving flexible options.
This is not a one-time decision. JetBrains will continue to evaluate new models as they become available to make sure they're recommending the right tool.
Keep JetBrains in your library
Save the videos and channels worth coming back to, and find them again in one place.
