AI enthusiast spent $2,618 on a single prompt and got Rome of 320 AD overnight

Pyotr Gostev from the Arena rating platform posted a video testing Claude Fable 5.2, walking through dozens of prompts. One of them cost $2,618 at Anthropic API rates.

Fable 5.2 ran at its maximum reasoning level (Max) on that request all night long. The result is a browser-based 3D reconstruction of Rome in 320 AD, the era of Constantine the Great.

Gostev asked the model to build a world where you can spend a day in the ancient city – a visitor appears above Rome, finds the Colosseum among the rooftops, and descends into any side street. In terms of quality, the result was supposed to rival an expensive historical game.

The spec specifically calls for the Sacred Way running from the Colosseum to the Forum, the imperial palace on the Palatine, and the Circus Maximus below it.

The model also got a test scenario – find the palace without any labels on the map, fly to the Pantheon, and linger in an ordinary neighborhood.

Gostev himself was measured in his assessment of the model:

Отличная модель, МНОГОЕ стало лучше, но не абсолютно все – Opus по-прежнему есть смысл использовать.

The full video with more examples:

Anthropic released Fable 5.2 in early November. The model costs $8 per million input tokens and $40 per million output tokens, twice the per-token price of Opus 5.5. For comparison, Anthropic engineer Felix Rieseberg built a shooter from a single prompt in eight hours, spending about $550.

Reactions to the demo were mixed. The level of detail in the city is impressive, but a closer look shows it's a long way from an "expensive game." Still, it works as a foundation for further improvements. The low frame rate also leaves a less than great impression, and the animations are pretty clunky.

Even so, for all the criticism, getting a result like this in a single night is very impressive.

More news
Tags: