Claude Opus 5 built a playable arcade game from one prompt; GPT-5.6 replicated it in Codex for $5

A community experiment is making the rounds on social media: one developer used Anthropic's Claude Opus 5 with a single prompt to generate a fully playable retro arcade boat racing game in one shot, burning roughly 690 million tokens and about $423 in the process.

The game, named INK TIDE, features an American cartoon style with complete countdown, lap, and ranking systems, a unified street-racing dashboard UI, and convincing boat pitch and roll physics. Developer Vyom said it was built with Vite, TypeScript, and Three.js using no external assets, with every model, texture, and sound generated by code.

The single prompt turns out to be a roughly 2,000-character requirements document: hard constraints first, then visual standards, functional requirements, development process, and delivery criteria, with the water system as its own chapter. It also specifies a detailed sub-agent division of labor, with separate agents for water, cartoon rendering, boat physics, AI opponents, character animation, audiovisual, and performance to fight attention dilution in large tasks, plus a screenshot harness and a visual-review sub-agent acting as a QA reviewer.

The author also revealed that the three AI opponents were deliberately weak: the fastest, KAIRA, is only 0.5 percent faster than the theoretical best line, and another opponent, PIP, makes two to three mistakes per lap — an obvious handicap likely meant to lower the barrier to entry.

The replica followed quickly: developer Anul Agarwal used Codex with an agent team, with GPT-5.6 Sol Ultra as the main agent handling orchestration and GPT-5.6 Luna Max as a sub-agent doing the development work, entering two prompts over about five hours at a cost of roughly $5, mostly converted from subscription quota rather than a strict API bill.

The replica keeps the original's most visible elements — driving a speedboat on the sea, cartoon-rendered water, a green track, racing UI, and multiple boats — but its water detail, character rendering, and visual consistency are noticeably rougher, putting the two builds at different levels of polish. Agarwal said GPT-5.6 Sol is very well suited to making games, and the remake used only about 20 percent of his weekly quota.

The takeaway of this head-to-head is that AI can now produce engineering-sound games, whether generated from a single prompt or cloned by a multi-agent team; the ceiling on final quality still depends on human judgment — and on your bill.