JetBrains has an advantage in this category that nobody else has, and it is worth stating clearly before the criticism. Their IDEs have spent two decades building a real semantic model of your code. Not text matching, an actual index that knows what this symbol refers to, where it is used, what type it resolves to, and what will break if you rename it. That machinery drives their refactoring tools, which are the best in the business, and it is available to their AI features in a way that a text based tool cannot replicate.
When the assistant proposes extracting a method or changing a signature, it is proposing something the IDE can verify. Suggestions arrive already knowing the types, and the completion is aware of what is actually in scope rather than guessing from nearby text. In practice that means fewer confident hallucinated method names, which is the single most tiring failure mode of code assistants. The local completion model is a good decision.
Short, everyday completions run on device with no network round trip and no metered cost, which keeps the common case fast and free while the expensive models handle the requests that genuinely need them. More vendors should do this. Consistency across the family matters if you work in several languages. The same features behave the same way whether you are in the Java, Python, or web focused IDE, which sounds mundane and saves real friction for people who switch daily.
And the escape hatches are welcome. You can supply your own API keys, and you can point it at a local model, which is meaningful for anyone in an environment where sending code to a third party is not permitted. Now the credit system, which is where this falls apart and which dominates every discussion of the product. Usage is metered in credits.
Each tier comes with an allowance. Requests consume credits at rates that depend on the model, the context and the operation, and the agent consumes them substantially faster than the chat does. The consequence is a mode of use I have not experienced with any competitor, which is hesitating before asking. You have a question, you know the assistant could answer it, and you weigh whether it is worth the credits.
That is a terrible experience to build into a developer tool, because the whole value of an assistant is that you reach for it without thinking. A subscription that makes you ration your own tool has misunderstood what it is selling. The agent makes this worse. Junie is a proper coding agent that plans, edits multiple files and iterates, and multi step agentic work is expensive by nature.
People on the higher tier report chewing through a large fraction of a monthly allowance in a few days of moderate use. So the flagship feature is the one you can least afford to use, which is an unfortunate place to end up. Predicting cost is nearly impossible. Before running a task there is no reliable way to know whether it will cost a little or a lot, because that depends on how many steps the agent takes and how much context it pulls, neither of which you control.
Competitors offering a flat rate for heavy use have simply removed this entire category of worry, and the contrast is not flattering. On quality, the agent is decent and it is not leading. The dedicated AI editors and the terminal based agents have been ahead on the harder end of agentic coding for a while, and the gap is visible on large multi file changes. Where this product wins is the tight, well grounded assistance inside an editor that already understands your project.
Where it loses is the ambitious autonomous task. Feature parity across IDEs is also patchier than the marketing implies, and working out which capability is available in which product at which version involves more searching than it should. My three point two is for a well engineered assistant with a genuine structural advantage in code understanding, undermined by a pricing model that penalises the use it is meant to encourage, by an agent that is both the headline feature and the most expensive one, and by a competitive position where the alternatives are cheaper to use heavily and currently better at the hardest tasks. If they moved to a flat rate tomorrow I would rate this considerably higher, and the code intelligence really is the best of the bunch.