Python 3.15 turns startup cost into a setting, not a property of your import graph

What lazy imports remove from startup is the module body, not the disk read. An import statement executes the target module in full at the point where it appears, which is how a tool that needs one function from a large library ends up paying for everything that library does while it loads, according to the dev.to account of the release candidate ^[3]. Deferring that execution until the code actually reaches a function or class inside the module ^[4] does not make the work cheaper. It makes the work conditional on the path the process takes.

The other headline items are not free in the same way. The two uses the write-up gives for frozendict, putting a mapping in another dictionary as a key ^[13] and stopping shared state from being changed by accident in threaded code ^[14], both require somebody to type frozendict where dict used to be, so nothing improves in code nobody edits. The free-threaded build removes the GIL in specific configurations ^[10], but the same account has NumPy still being reworked to exploit it, with hidden bottlenecks still being found ^[11]. That benefit arrives on someone else's release schedule.

Deferral also has a cost the write-up does not mention. Whatever a module does while it executes, including raising, now happens the first time the program reaches into it rather than during startup ^[4]. For a short-lived command the distinction is academic. For a long-running service, an import moves out of boot and into whichever request touches it first, which is a worse place to discover a missing dependency or a slow module-level initialisation.

Worth being blunt about the evidence. This is a single write-up that names no PEP and publishes no benchmark ^[3], and 3.15 is still a release candidate ^[1], so treat the list as provisional rather than as shipped behaviour. The same piece reports a fix to a corner case in the lower-case conversion method that could return inconsistent results and weaken logic built on string normalisation ^[15], but gives no issue or advisory identifier, which is precisely what you would need to work out whether your current deployment is exposed. The JIT is described as receiving further refinement toward automatic speedups that avoid rewriting logic in a lower-level language ^[9], again with no figures attached. Tachyon is the one item you can assess without waiting for anything, because sampling a live process needs no restart and no instrumented build ^[8], so you can point it at a slow service you already run and see whether its picture of the stack is good enough to act on.