Astra Overwhelms Fable in Computer Usage Capability

The AI's drawing ability was measured by replicating and drawing the same 'thumbs-up' image.

For the test, the AI was instructed to use its native computer use function. Conditions were set to replicate an attached image using Krita's paint tools and to stop working upon achieving over 90% similarity. Additionally, it was clearly instructed that the reference image should be displayed in Krita, while the actual work should be done on a blank canvas.

Astra showed a systematic working method, first establishing the overall shape before proceeding with detailed drawing. After reviewing the screen once, it would refine the shape, select appropriate brushes and thicknesses for the situation, and repeat the process of modification (overpainting/deletion). During the work, Astra employed delicate strategies such as zooming in on the canvas to enhance precision.

In this environment, Astra performed the task for a total of 49 minutes.

However, it received positive evaluations for hair texture, color, and shade expression. The entire task took 49 minutes.

Fable initially seemed to attempt texture expression but gradually adopted a rougher working style. It used the largest brush to crudely fill in shapes, which contrasted with Astra's delicate approach.

Astra meticulously switched tools while working, whereas Fable took a long time to deliberate and mainly used thick brushes. Comparing click feedback and tool usage frequency, Astra was observed to use far more tools faster and more accurately.

The initial work results fell short of realistic quality, and overall, the completeness was low, with rough expressions and fewer strokes being prominent.

Astra extracted colors and worked precisely by switching between photos and the canvas, while Fable's work completeness was significantly lower. In particular, Astra's fast interface understanding and command input speed enhanced its efficiency.

In the comparison of tasks using Mac Numbers files, Astra opened files very smoothly. Fable, on the other hand, repeatedly encountered errors during file retrieval and navigation, struggling from the start. This showed a significant difference between the two models in the stability of system call stages when performing computer usage missions.

In data calculation and visualization tasks within Numbers sheets, both Astra and Fable performed accurately, yielding identical final numerical results. However, there was a significant difference in time taken. Astra completed the task in 13 minutes and 50 seconds, while Fable took over 2 minutes just to open the file, showing significantly slower overall processing speed.

Astra, in contrast, can perform system control tasks relatively freely with just prompts, offering the advantage of less user intervention.

This clearly demonstrates that Astra processed tasks overwhelmingly faster than Fable.

In the Keynote creation performance test, Astra completed the task in 16 minutes and 3 seconds, showing precise execution without misclicks. Fable, on the other hand, took 29 minutes and 49 seconds, experiencing frequent misclicks and interface errors, leading to many trial-and-error attempts. The host explained that despite providing specific rules for slide composition during Keynote creation, Fable's errors were prominent.

Drawing was analyzed as a highly difficult task requiring delicate judgment and control over brush thickness, movement, and XY coordinates. Astra demonstrated superiority in these delicate brush strokes and control abilities, while Fable was found to lack the ability to perform delicate tasks.

This suggests that the user environment should be considered based on the nature of the task and whether the AI model's fast mode is used.