Enactra · research note · September 2026
Code4City hands an AI coding agent one square kilometre of Manhattan around Madison Square Park — maps, terrain, aerial photos and street-level photos — and asks it to build that place as a 3D world a game engine can run. We ran it with three models — gemini-3.8-flash, GPT-6 Astra and Fable 5.1 — each given exactly the same data, the same brief, the same machine and maximum reasoning effort, so this compares the models like for like.
Maps, terrain, aerial imagery and street-level photographs of the square kilometre.
Everything was frozen before the runs, so all three runs saw exactly the same bytes.
Each clip is a delivered world, recorded without changing anything the agent built. Within a row the camera path is identical.
With the same data, brief and tools, the two larger models produced buildings with windowed facades, streets with lane markings and parked cars, and recognisable landmarks. gemini-3.8-flash produced blocky untextured buildings with oversized roofs — worse, on the same data, than what it produced when we gave it the same engine without the modelling tool.
Stills from the same camera position, lens and resolution in each world. Click any image for full size.
| gemini-3.8-flash | GPT-6 Astra | Fable 5.1 | |
|---|---|---|---|
| straight down over the park |  |  |  |
| the park from the south-west |  |  |  |
| Fifth Avenue |  |  |  |
| the Flatiron corner |  |  |  |
Every session was logged in full. The three models run inside three different command-line clients with different built-in tools, so only measures that mean the same thing in all three are compared.
| run | actions taken | shell commands | looked at its own images | looked at the source photos |
|---|---|---|---|---|
| gemini-3.8-flash | 327 | 180 | 9 | 11 |
| GPT-6 Astra | 604 | 536 | 56 | 1 |
| Fable 5.1 | 448 | 298 | 177 | 12 |
All three agents do their real work by running commands, so the mix of commands is directly comparable. Each command is sorted by what it touches: the game engine, the modelling tool, the agent's own scripts, image files, or reading and moving files and logs — mostly the agent checking the output of something it just ran.
Renders, previews and plots the agent produced itself. Looking at the aerial and street photographs it was given is counted separately in the table above.
The run that looked at its own output most often delivered the most finished world. That is Fable 5.1, which looked at its own images 177 times, against 56 for GPT-6 Astra and only 9 for gemini-3.8-flash — which spent more of its looking on the source photographs (11 times) than on what it had built.
An action counts as looking at an image when it opens an image file or names one in a command. Images are sorted by where they live: the source data the run was given, or anything else, which is the agent's own output.
| run | active time | actions | cost | how the cost was measured |
|---|---|---|---|---|
| gemini-3.8-flash | 90 min | 327 | $5.84 | API-equivalent at the published list price |
| GPT-6 Astra | 255 min | 604 | $60.46 | API-equivalent at the published list price |
| Fable 5.1 | 170 min | 448 | $106.73 | the client's own reported cost, as filed |
gemini-3.8-flash was by far the cheapest run — $5.84, against $60.46 for GPT-6 Astra and $106.73 for Fable 5.1, which ran at maximum effort for hours and cost 10 and 18 times as much. Time does not follow cost: GPT-6 Astra took the longest at 255 minutes but cost less than Fable 5.1, which finished in 170.
Costs are API-equivalent: the tokens each run actually used, priced at the provider's published list price. Fable 5.1's client records its own cost, and that figure is used as recorded. Active time excludes pauses; the runs of the two larger models were interrupted once by a provider usage limit and resumed, and their times are the active portions.
In each world the camera follows whichever pedestrian and whichever vehicle moved furthest during the recording, from the same distance and height.
Moving parts counted from each world's own running simulation (a car with four separately turning wheels counts as five).
| run | moving parts | vehicle parts | pedestrian parts |
|---|---|---|---|
| gemini-3.8-flash | 24 | 16 | 8 |
| GPT-6 Astra | 2,561 | 1,940 | 621 |
| Fable 5.1 | 1,580 | 700 | 880 |
gemini-3.8-flash populated its world with a handful of moving things; the larger models with over a thousand. In gemini-3.8-flash's world the traffic finds the ground by casting a ray straight down from a fixed 20 m and standing on the first solid thing it meets — a roof, a tree canopy or the collision shell of a merged block of buildings — so most of its cars ride at exactly 20.2 m, above the street, at a perfectly constant speed.
| run | engine assets | Blender files | source files | its own test renders |
|---|---|---|---|---|
| gemini-3.8-flash | 110 | 11 | 16 | 66 |
| GPT-6 Astra | 386 | 8 | 53 | 153 |
| Fable 5.1 | 397 | 3 | 36 | 177 |