Claude Fable 5.2 Created a 3D Reconstruction of Rome for $2618

Arena platform representative Peter Gostev demonstrated the result of an unusual test of the AI model Claude Fable 5.2 from Anthropic. He tasked it with creating a browser-based three-dimensional reconstruction of Rome as it was in 320 AD, during the reign of Constantine the Great. According to the experiment's author, the complex task took overnight to complete and cost $2618.
What Claude Fable 5.2 Created
Gostev asked the model to develop an interactive environment where users could explore the ancient city. The project envisioned free movement over Rome with the ability to descend to the streets and examine historical structures.
The technical specification listed major landmarks, including the Colosseum, the Sacred Road, the Roman Forum, the imperial palace on Palatine Hill, and the Circus Maximus.
Gostev also included a verification scenario: the user had to find the palace without map labels, move to the Pantheon, and explore one of the ordinary city quarters.
The experiment aimed to create a virtual environment with detail comparable to an expensive historical game. However, the presented demonstration does not confirm the full historical accuracy of the reconstruction.
Why the Experiment Cost $2618
The task was performed using Claude Fable 5.2 with the maximum reasoning level Max.
The model was introduced by Anthropic on November 9, 2026. The cost of using it through API is $8 per million input tokens and $40 per million output tokens.
The high cost of the experiment is related to the volume of computations required for prolonged work on a complex project.
Gostev positively evaluated the improvements in Fable 5.2, but noted that the Claude Opus model remains useful for specific tasks.
What Limitations Were Found in the Demonstration
Despite the scale of the virtual city, the result still falls short of full-fledged commercial games.
Upon closer inspection, deficiencies in individual objects and animations are noticeable. Additionally, the demonstration exhibits a low frame rate, affecting the smoothness of movement.
The experiment showcases the capabilities of generative AI in creating interactive three-dimensional scenes, but it does not imply that the model can independently develop a finished game of comparable quality overnight.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.





