XNews

Model Recommendations and Demos (Cursor & Gemini)

Model Recommendations and Demos (Cursor & Gemini)

Cursor published mode-specific model recommendations and Gemini demo compared versions. - Cursor recommended gpt-5.2 extra high for planning mode - Cursor recommended gpt-5.2 extra high for debugging mode - Cursor recommended Claude Opus 4.5 and Composer-1 for agent mode - Cursor recommended Composer-1 for Q&A mode - The recommendations aim to balance intelligence, speed, and cost - Gemini 3 Pro was shown playing Pokémon against Gemini 2.5 Pro - The demo noted prior impressions of 2.5 Pro and highlighted 3 Pro improvements - Visual/video comparison accompanied the Gemini demo

Key insights

Cursor maps specific models to operational modes

Cursor team member Eric shared explicit model recommendations per Cursor mode: planning and debugging modes should use gpt-5.2 extra high; agent (smart-agent) mode should use Claude Opus 4.5 and Composer-1; question-and-answer mode should use Composer-1. The mapping is presented as a practical combination balancing intelligence, speed, and cost constraints.

Gemini 3 Pro demonstrated improved gameplay versus 2.5 Pro

A demonstration compared Gemini 3 Pro playing Pokémon against Gemini 2.5 Pro, noting that observers were previously impressed by 2.5 Pro and that 3 Pro shows further improvement in the same task. The post included a visual/video comparison of the two model versions.

Source posts