JetBrains logo, a pink and orange square mark beside the word JetBrains, for a guide to Junie Eco Mode

Junie Eco Mode: Power-User Guide to Eco, Junie Lite and Junie Local

by

Last updated: November 1, 2026

Junie Eco Mode, released in October 2026, splits each coding task between two models. Your chosen model reads the code and writes a plan, then a cheaper model, currently GPT-6 Luna, edits files and runs tests. JetBrains reports roughly half the model cost with similar resolve rates. It is generally available in Junie CLI 26.10.20.

Contents

What’s new in Junie in October 2026

Junie is the AI coding agent from JetBrains. It runs in JetBrains Air inside JetBrains IDEs and as the Junie CLI in a terminal. October brought four notable changes, documented on the Junie CLI What’s new page and the Junie blog.

Change Date What it does
Junie Lite October 2, 2026 Free model tier for every JetBrains account in Air and Junie CLI
Junie Local on RTX Spark and stable Windows October 16, 2026 Local agent on NVIDIA RTX Spark PCs and 24 GB RTX GPUs
Eco mode (CLI 26.10.20) October 21, 2026 Plan with your model, build with a cheaper model; default model selection; quota in /account
CLI 26.10.27 October 28, 2026 GPT 6.1 Terra support
Junie Local on Mac (Splash and DFlash2) October 2026 About 70 tokens per second on M5 Max, up from about 40; now runs on M4 Macs with 64 GB

Earlier October CLI builds added Figma MCP support (October 15), Gemini 4 Argon (October 16), automatic cache cleanup with a new junie cache [clean] command, and per-session model, effort and brave settings (October 13).

How Junie Eco Mode works

According to the JetBrains Eco mode announcement, Eco mode treats the model you pick as the architect and a cheaper model as the builder. There is one handoff per task, and the work never switches back and forth.

Step 1, the plan. Your selected model, at your selected effort level, investigates the task. It can read and search code but cannot change anything, except that it may write a small test that reproduces a reported bug. It usually takes about 15 steps and produces a plan with a fixed structure: goal, scope, where to change, approach, tests, and risks and corner cases.

Step 2, the build. Junie passes the same conversation and the plan to a much cheaper model, currently GPT-6 Luna. That model edits files, runs the build and tests, and fixes failures.

When Eco stays off. Planning runs only for code changes you intend to keep. Questions, explanations, project setup and “run this and tell me what happens” requests stay with your chosen model. Eco mode is unavailable in /remote sessions and with some model and effort combinations, including Opus 5.5 at low effort and Sonnet 5.5.

Default-on models. In supported environments, Eco is on by default for Opus 5, Opus 5.5 (medium effort and above), Fable 5.1, Grok 4.5, 4.6 and 4.7, GPT-5.6 Sol, GPT-6 Sol, GPT-5.6 Terra and GPT-6 Astra.

Eco mode cost and resolve-rate results

JetBrains tested each model and effort configuration with Eco off and on across two internal benchmarks of real tasks from private repositories. On benchmark B, average model cost per task fell from $2.61 to $1.11, while the average resolve rate moved from 49.1% to 48.5%. On benchmark A, cost fell from $2.51 to $1.30, and resolve rates rounded to 32.1% either way. A separate run on 100 benchmark B tasks gave these examples:

Setup Resolve rate Model cost per task
Opus 5.5 (max) 55% $4.91
Opus 5.5 (max) + Eco 56% $1.81
Fable 5.1 (high) 53% $3.17
Fable 5.1 (high) + Eco 52% $1.38
Grok 4.7 (high) 46% $3.37
Grok 4.7 (high) + Eco 52% $1.70
GPT-5.6 Sol (high) 43% $1.16
GPT-5.6 Sol (high) + Eco 42% $0.58

These are vendor benchmarks, and JetBrains notes that individual configurations can show larger changes in resolve rate.

How to use Junie Eco Mode

  1. Update Junie CLI with junie update so you are on 26.10.20 or later, or use JetBrains Air in a 2026.2 JetBrains IDE or the IntelliJ IDEA 2026.3 Beta.
  2. Run /model and choose a supported model and effort level, such as Opus 5.5 at medium effort or higher.
  3. Confirm Eco is switched on in the model picker. It is on by default for supported combinations.
  4. Describe a code change you want to keep. Junie plans with your model, then hands the plan to the builder model.
  5. Review the diff and test results as usual. Turn Eco off in the model picker for tasks where you want one model end to end.
  6. Run /account to see remaining quota and your active license.

Junie Lite: the free tier

The Junie Lite announcement describes a free tier included with every JetBrains account, including free accounts and AI Pro and AI Ultimate subscribers. Usage is “virtually unlimited,” inference runs only on JetBrains infrastructure, and prompts and code are not used for training. The model is Qwen 3.6 27B today and may change. JetBrains says the offer is valid until December 31, 2026. In the CLI, run /model to select it.

Junie Local: faster and on Windows

Junie Local runs the agent entirely on your machine at no token cost. On Mac, the Splash and DFlash2 update replaced the MLX-VLM engine with Splash and added a DFlash2 draft model. JetBrains reports generation speed rising from about 40 to about 70 tokens per second on M5 Max, prefill rising from 850 to 1,200 tokens per second, and about 24 GB of RAM at 100k context. M4 Macs with 64 GB of RAM are now supported, though prefill is about four times slower than on M5.

On Windows, the RTX Spark announcement made support stable for NVIDIA RTX Spark PCs and x64 Windows 11 PCs with RTX GPUs that have at least 24 GB of VRAM. You need NVIDIA driver 581 or newer and about 20 GB of free disk space. You can choose Qwen3.6-27B or the Qwen3.8-3.6-27B-blend model. On Mac, run junie update, then /local, then press Ctrl+R to update the inference engine.

Practical workflows and limitations

  • Tiered work: plan a migration with a frontier model plus Eco, then switch mechanical follow-up edits to Junie Lite.
  • Private code: use Junie Local when code must not leave the machine, and accept slower prefill on M4.
  • Hard investigations: turn Eco off for open-ended debugging where you want your strongest model to keep control.

Limitations: JetBrains has not published per-token prices for Eco mode, and IDE support for other 2026.3 Beta builds is still rolling out. For comparable approval and cost controls in other agents, see our guides to Kiro auto approve, Devin Desktop 3.11, Codex CLI 0.166 and Claude Code new features. For other local model runners, see our Ollama October updates guide.

Frequently asked questions

What is Junie Eco Mode?

Eco mode is a Junie setting from JetBrains that lets your chosen model investigate a coding task and write a plan, then hands implementation and testing to a cheaper model, currently GPT-6 Luna.

How much does Junie Eco Mode save?

In two JetBrains internal benchmarks, Eco mode cut average model cost per task by 57% and 48%, while average resolve rates changed by less than one percentage point. Results vary by model and workload.

Which Junie CLI version has Eco mode?

Eco mode is generally available in Junie CLI 26.10.20 and later. That release, dated October 21, 2026, also added default model selection and remaining quota in the /account command.

How do I turn Eco mode off in Junie?

Use the model picker. Eco mode is on by default for supported model and effort combinations, and you can switch it off when you want one model to handle the whole task.

Is Junie Lite free?

Yes. JetBrains says Junie Lite is a free tier included with every JetBrains account, with virtually unlimited usage. It currently runs Qwen 3.6 27B, and the offer is valid until December 31, 2026.

What hardware does Junie Local need?

Junie Local runs on M4 or newer Macs with 64 GB of RAM, NVIDIA RTX Spark PCs, and x64 Windows 11 PCs with an NVIDIA RTX GPU that has at least 24 GB of VRAM.

Sources

Image: JetBrains logo by JetBrains, public domain (trademarked), via Wikimedia Commons. Placed on a white background and resized. FreeAnswers is not affiliated with JetBrains.



Related answers

Leave a Comment