Same job, three models: build a racing kart game for the browser, then make it playable on a phone.
All three pass the same 10 basic checks (3 laps, bots finish, results screen, restart...). The extras are where they differ.
Haiku is about 200x cheaper and ready in minutes, but needs a fix and has far fewer extras. Opus costs the most and takes the longest.
Opus ran at higher effort and Haiku got a one-line prompt, so the gaps mix model, effort and prompt. Games were tested headlessly, not played by hand. Cost and time come from the Claude Code session logs (idle gaps removed); one run per model.