Gemini 3.0 Is Finally on Cursor? (New Stealth Model)

Cheetah Model Overview and Initial Testing

Introduction to Cheetah Model

• A new model named Cheetah is available on Cursor, priced at $1.25 per million input tokens and $10 per million output tokens, similar to Gemini 2.5 Pro.

• Speculation exists that Cheetah may be a variation of Gemini 3, as Google DeepMind's developer lead hinted at upcoming developments.

Context of AI Model Releases

• The release cycle of AI models is discussed, noting that OpenAI has recently launched GPT-5 and Claude has released updates like Claude Sonnet 4.5.

• Anticipation builds for Google's next move in the AI space, with expectations for Gemini 3's arrival.

Testing Cheetah on Real World Tasks

Refactoring Anki Add-on

• The speaker plans to use Cheetah to refactor an Anki add-on called AI Language Explainer, which was initially built using Opus 3.

• The task involves reorganizing the code into multiple files while ensuring functionality remains intact.

Performance Observations

• Initial tests show that Cheetah processes files quickly; it appears faster than previous versions like Gemini 2.5 Pro.

• After refactoring, the add-on does not crash upon launch but encounters issues with visibility in the edit section.

Challenges Encountered During Testing

Debugging and Adjustments

• The model's default behavior includes adding debugging logic, which is seen as a positive feature despite initial errors encountered during testing.

• After several attempts and adjustments based on feedback from the model, functionality improves significantly.

Comparative Analysis with Other Models

Performance Comparison

• When compared to GPT-5 in Codex CLI, GPT-5 performed better in similar tasks due to potentially optimized system instructions for its environment.

• There are doubts about whether Cheetah is indeed Gemini 3 based on performance metrics observed during testing.

Future Development Plans for HyperWhisper Application

New Features Implementation

• The speaker aims to implement a statistics page within the HyperWhisper application that tracks transcription metrics such as time spent and word count.

Transcription Efficiency and Image Generation with AI

Overview of Transcription Statistics

• The speaker discusses a statistical graph that tracks daily transcription metrics, including minutes transcribed, words transcribed each day, and total words. It highlights the efficiency gained from using voice-to-text technology.

• The transcription process is noted to be significantly faster than traditional writing methods, suggesting a potential time-saving calculation based on the difference between spoken and written word counts.

Project Development with AI Models

• The speaker expresses interest in creating a new project utilizing Google's Nano Banana image model within a Next.js application for generating variations of uploaded images.

• Features desired include customizable prompts for poses and an interactive canvas for drawing poses to be processed by the AI model.

Technical Challenges Encountered

• Initial attempts to run the application reveal issues with API key exposure in client-side code, indicating poor security practices.

• Errors arise when trying to generate image variations due to incorrect documentation being referenced; this highlights challenges faced during development.

Model Performance Observations

• Despite encountering errors, the model successfully generates images after multiple attempts. However, some outputs are unexpected (e.g., added skeleton), raising questions about the model's reliability.

• The speed of processing leads to speculation about whether it is an advanced version of existing models like Gemini 3 or another variant designed for rapid output without deep reasoning capabilities.

Reflections on Model Usability

• The speaker concludes that while they may not use this model for large tasks, its efficiency makes it suitable for smaller projects where quick iterations are beneficial.

• Emphasizing the importance of fast models in maintaining workflow continuity, they express a preference for tools that minimize context switching during coding sessions.

Invitation to Join AI Startup Community