The headline is the wrong unit.
Piotr Migdał gave two coding agents the same short brief. Each had to build a Three.js visualization of all 55 cities in Italo Calvino's Invisible Cities without follow-up questions. The published account says the Opus 5.5 run stopped after one hour and 25 minutes, used six subagents, and cost about $74 in API tokens. The GPT-6 Astra run took 53 minutes and cost about $10.[1]
Those numbers describe a run. They do not score the result. More elapsed time would not establish better taste. A lower bill would not establish better interpretation. The article's strongest evidence is the pair of inspectable sites and repositories, not the dramatic time limit.
Speed earns a look. Inspection decides whether the work deserves a release.
The artifacts are real and different.
The Opus repository contains 55 coded dioramas, three board layouts, a tour, keyboard controls, optional sound, and deep links. Its README says the text is paraphrased for the atlas.[2] The Astra repository also contains 55 miniatures, thematic trails, a city index, journal state, optional sound, postcards, and a non-WebGL fallback. Its README documents a Pages test for the deployed subdirectory.[3]
Both public sites returned HTTP 200 when checked. We opened the Opus site and confirmed that its entry screen and 55 selectable canvas groups loaded. We did not audit all city models, compare every paraphrase with the book, or certify mobile and accessibility behavior.
Output is not process evidence.
A repository preserves the result and some operating instructions. It does not automatically preserve the exact brief, every agent action, the state inspected before a correction, or the reason a visual choice survived. Without those records, a reviewer can inspect the current artifact but cannot reconstruct the creative path.
The Hacker News discussion split around that missing path. Some readers treated the sites as striking technical sketches. Others found broken geometry, weak interpretation, or little reason to inspect details.[4] The thread is useful as a map of reader reactions. It does not verify how either project was built.
Keep five records.
Save the brief with constraints and allowed tools. Record ordered actions. Name the exact revision or image the agent inspected. Tie each revision note to an observed problem. Then replay the saved work from a known revision and compare the result.
A successful replay proves only that the retained route can rebuild a comparable artifact. It does not prove originality, authorship, taste, or release fitness. A human review still has to decide what the work means, what it gets wrong, and whether it should ship.