GPT-6 Astra Stumbles on Viral Canva Test

This is a useful reality check, not a definitive model benchmark: computer-use demos can be highly sensitive to prompts, reference images, app state, retries, and editing.

• –Astra’s achievement is GUI control—placing shapes and strokes in Canva—not native image generation
• –A failed replication does not disprove the original demo, but weakens claims that the workflow is reliably push-button
• –Fair comparisons require identical prompts, source images, browser state, permissions, time budgets, and retry limits
• –Developers should evaluate full action trajectories and intermediate failures, not only polished final clips
• –Strong desktop-use benchmark scores do not guarantee consistent creative performance in consumer design apps