Seven beliefs about developer tooling that quietly increase what you spend, on infrastructure or on the far more expensive resource of engineering time.
The reality: the free options cover most work. VS Code, Neovim and Helix all speak the Language Server Protocol, so completion, go-to-definition and diagnostics come from the same language servers a paid IDE uses. Paid IDEs earn their licence on deep refactoring in large typed codebases and on integrated debuggers for JVM and .NET work. If you are writing TypeScript and Python, you are paying for capability you rarely reach.
The reality: every job carries fixed overhead — spin-up, checkout, dependency install, container pull. Split a four-minute suite into eight jobs and you may pay eight times the setup cost to save two minutes of test time. Measure the overhead per job first; often the bigger win is caching dependencies properly, or running on one larger machine where the setup happens once.
The reality: only if you pinned everything. FROM python:3.12 resolves to a different image next month, apt-get install pulls whatever is current, and an unpinned pip install does the same. Pin base images by digest, use a lockfile, and your builds become reproducible. Without that, a container is a consistent environment, not a reproducible one.
The reality: the report does not distinguish reachable from unreachable. A regular-expression denial-of-service in a build-time dependency that only ever processes your own file paths is not an exploitable production risk. Triage by whether the vulnerable code path runs with untrusted input at runtime. Treating every finding as urgent produces forced major upgrades, breakage, and eventually a team that ignores the tool entirely.
The reality: it is cheaper for spiky and low-volume workloads, where you genuinely pay nothing while idle. There is a crossover point: at steady, high request rates, per-invocation pricing exceeds the cost of a right-sized always-on instance. Work out your own crossover from real request volume rather than assuming either direction.
The reality: you swapped a per-minute bill for a per-hour one, plus maintenance. A runner VM costs money while idle, needs patching, and needs someone on call when the disk fills with Docker layers. It pays off with sustained high utilisation or when you need specific hardware. For bursty workloads, hosted minutes are usually the cheaper line item once you count the engineer.
The reality: billing is per minute times a size multiplier, so a machine that costs twice as much and halves the wall-clock time is roughly cost-neutral — and it returns developer minutes for free. Time one representative job on each available size before assuming the smallest is the cheapest.