Every model comparison starts the same way and ends in the same argument: someone quotes a leaderboard number, someone else says it does not match their experience, and nothing is decided. This playbook is the version that produces a decision, using four flagships from this directory as the running example.
Step one: write the job as a prompt
Not a category of job; one prompt with real input. "Summarize this incident report into three bullets for an on-call handoff" is a prompt. "Handle incidents better" is a mood. Copy the prompt and the input into a file so every model sees byte-identical text.
Step two: grade without the model names
Run the prompt through each candidate, paste the outputs into one document with the names stripped, and have someone who knows the job rank them before they learn which model produced which. This is the only step that reliably beats a leaderboard, and it is the step people skip.
Step three: discard on constraints before scoring on quality
A model whose license you cannot accept, whose context window cannot hold the input, or whose price makes the call unaffordable is out before the ranking matters. The comparison table on this site highlights the best cell per row for exactly this purpose: as a filter, not a verdict.
Step four: keep two, not one
Keep the best-quality pick and the cheapest acceptable pick. Route the cheap one to the high-volume classification and extraction calls and the expensive one to the work where a wrong answer costs more than the tokens. That split, more than any single choice, is what makes a model decision hold up over a quarter.
The compare page builds the table from whichever listings you select, and the written verdict underneath uses only your rows. The deciding after a model comparison guide covers what to do with the result.