All three verdicts in the account are statements about context acquisition, not about model quality. Copilot is described as most convenient for everyday autocomplete, boilerplate and small functions [4], and as good when the problem sits close to the code being edited [13]. Cursor is described as an AI-focused editor that made it easier to reference related files and adapt a function to its surroundings [3][10]. Claude Code is described as strongest when the task crossed several files or ran multiple steps from a terminal [5][15]. Read those as three answers to the same question: what the tool has in front of it before it writes. That reading is mine. The account does not describe how any of the three assembles its context, only what the author reports it could use [6].
Weigh the evidence before reorganising procurement around it. The setup lists five kinds of work: writing code, debugging errors, refactoring functions, creating tests, and changes across multiple files [1]. The text supplied reports per-tool outcomes for two of those, implementation and debugging [10][13], and it ends mid-sentence in the refactoring section [18], which leaves three of five task types without a reported result [1]. There are no timings, no pass rates, no token counts and no prices in it [17]. One developer, one month and no stopwatch is a hypothesis with a mechanism attached, not a benchmark.
For the ranking to transfer, several things have to be true of your work. You have to be editing an existing codebase rather than starting one, which is why the author refused the usual build-a-todo-app prompt [7]. The worked examples are JavaScript and React, going by the async function signature and the UserList.jsx stack frame [2]. Your repository has to hold prior implementations worth imitating, which is what makes "inspect how similar functions were done elsewhere first" a useful instruction rather than a delay [10]. And your bugs have to surface somewhere other than where they originate [15]. If you write new single-file code all day, the ranking flattens: all three produced a working starting point on the missing-function task, and the difference only appeared during cleanup [9].
The cost of picking one is specific. Choose buffer-scoped completion for a codebase where the cause sits three files from the crash, and you become the context assembler, handing over the component, the API function and the data structure yourself [13]. That labour does not vanish. It moves to the person with the keyboard. The technique that survives whatever you buy is prompt shape: supplying the related files and asking for the root cause produced better answers than asking for a fix [12]. If I were running this test to justify a second subscription, I would log one number per task, the count of times I had to hand a tool a file it should have found on its own. That is the number the licence is actually buying down.
