GPT 6.1 Terra
After Haiku 5.5, it is GPT 6.1 Terra's turn to paint a portrait in code. As in the original experiment, the brush is a translucent ellipse, the canvas is black, and every stroke is a shade of grey.
One shot
Terra wrote this portrait in one pass without seeing a render. It used 45 strokes out of the 300 available.
The eyes catch the light and the face is immediately legible, but the soft edges make the portrait feel ghostly. The long neck seems to float above the shoulders, and the collar is reduced to two sharp lines.
A study in stages
The second portrait uses a simple frontal proportion guide: eyes halfway down the head, nose at three quarters and mouth at seven eighths. The sequence moves from large shapes to facial planes, features, hair and clothing, with the same growing stroke budgets as the Haiku study.
This is a staged construction, not the same canonical test: the stages were prepared together and rendered afterwards, without inspecting and revising each one before the next. It is therefore not a comparison of the models' ability to use visual feedback.
The early planes suggest a face before the features arrive. The finished study has a direct gaze, but shares the one-shot portrait's disconnected neck and sharp collar. Most of the late strokes are extremely faint marks added to reach the stage budgets, so the last few images change very little. A higher stroke count did not make for a more resolved picture.
Verdict
Terra made a recognizable portrait with surprisingly few strokes, but the later stages did not meaningfully improve it. If anything, the study shows how easily extra strokes can be spent without resolving the picture. When asked to assess the result, Terra could identify the problems: the disconnected neck and shoulders, the sharp collar and the faint late marks. Knowing what was wrong did not translate into a better painting here. Since the stages were prepared together, this run does not establish whether Terra could correct those problems through repeated visual feedback.