Google's Gemini 4 Pro Checkpoints Show Divergent Output Styles on Identical Creative Prompts

Multiple internal checkpoints of Google's unreleased Gemini 4 Pro language model are producing drastically different outputs when given identical creative prompts, raising questions about consistency as the company prepares for a public launch expected in October 2026. When presented with the task "SVG of a pelican riding a bicycle," different checkpoints generate outputs ranging from fully interactive, elaborate scenes complete with animation controls, lighting effects, and pedaling-speed adjustments—spanning thousands of lines of carefully coordinated SVG code—to minimal, clean interpretations of the same concept. The divergence has sparked discussion in the AI community about whether such variation represents a fundamental challenge with the model or reflects deliberate testing of different generation modes.