Notes on film, new tools and the decisions behind the work.

Tools / Analysis

Can GPT-6.2 make a city stand up?

Astra and Sol got the same city to animate and the same pelican to draw. Their published results help explain why a 3D leaderboard and a director can disagree.

A pelican riding a bicycle in front of the names Astra and Sol, a Signals illustration of Simon Willison’s drawing prompt.
DU / SIGNALS · S/120Signals illustration of Simon Willison’s pelican-on-a-bicycle prompt

Give Astra and Sol the same directing note. Take Yonkers, New York, use real OpenStreetMap data, and make the city rise into a giant walking machine. The hard part comes a beat before the first footstep. Can you still see the city in the thing that just stood up?

Watching Yonkers take its first step

Simon Willison gave that assignment to both GPT-6.2 models in Codex CLI at Ultra effort. He published both scenes. In his judgment, Claude Fable 5.2 and Gemini 4 Carbon had handled the transformation more convincingly. Then he gave Astra a second job. Take the two results, put them one above the other, label them, and extend the comparison from ten seconds to twenty. Open his write-up and you get to judge the very moment he was trying to describe.

Sources: Simon Willison [1] · Simon Willison [2] · Simon Willison [3]

VIDEO / CASE STUDY

Astra and Sol animate Yonkers

Credit / Simon WillisonOpen original 

Willison’s twenty-second comparison shows the two cities moving together. The full interactive scenes are linked below.

A pelican, two pedals and a 3D leaderboard

Willison also asked each model to draw a pelican riding a bicycle in SVG. At maximum effort, he calculated that Astra’s drawing had cost 88.71 US cents and Sol’s 13.69 cents. You can put the two drawings side by side and follow the bird’s legs to the pedals. That detail is easier to judge than a general promise of better drawing.

There’s a good reason to put that example next to the benchmark. In Artificial Analysis’s human-rated interactive 3D evaluation, Astra led with 1,318, ahead of Fable at 1,271 and Sol at 1,259. That gives me a reason to try Astra on a 3D brief. Willison gives me a criterion for deciding whether to keep its shot. Watch the street pattern as the body turns. If the audience loses the connection there, a beautifully rendered robot still misses the direction.

Sources: Simon Willison [1] · Artificial Analysis [4]

The question the agent sent back

Theo Browne found a different kind of friction during prerelease testing. Two hours into building an Outer Wilds level, the agent asked him to choose between two internally named video files. Willison reproduces the example. I’d want the agent to ask me how the level should play. Choosing which temporary clip to write is precisely the decision I thought I’d handed over.

Sources: Theo Browne / X [5] · Simon Willison [1]

Where Astra and Sol landed on the Agent Arena Index

On November 19 Arena published another comparison, this time in its index for AI agents. With both models set to Max, Astra ranked sixth on 1,291 points and Sol ranked eleventh on 1,241. Astra finished ahead of Sol on the index. Choosing a model to animate Yonkers still calls for a closer look at the city itself. Does its weight shift when it takes a step? The ranking gives you a reason to try Astra first. The clip gives you something to judge.

Sources: Arena [6]

SOURCES & METHOD

The record behind the article

Checked 19 November 2026. Technical claims are grounded in developer documentation; videos are credited to their creators. Company claims and third-party reviews are attributed in the text; conclusions are editorial analysis. No original comparative testing of these releases was conducted for this article.

  1. GPT-6.2 Astra and Sol, pelicans and Yonkers Simon Willison
  2. Yonkers scene generated with Astra Simon Willison
  3. Yonkers scene generated with Sol Simon Willison
  4. GPT-6.2 Astra and Sol evaluation results Artificial Analysis
  5. Prerelease coding experience Theo Browne / X
  6. GPT-6.2 Astra and Sol on the Agent Arena Index Arena · 2026-11-19