Brief Summary
This video compares Google's Gemini CLI with Cloud Code for developing a real-world production application, specifically enhancing the Tensor AI news app. The comparison focuses on ease of use, performance, and cost-effectiveness in implementing features like swipe navigation and audio generation. The initial free tier of Gemini CLI seems promising but hits roadblocks due to rate limiting, leading to an upgrade. Despite a larger context window, Gemini CLI requires more guidance and produces less satisfactory results compared to Cloud Code in this specific example.
- Gemini CLI is initially free but performance degrades with rate limiting.
- Cloud Code provides better solutions with less guidance in the tested scenario.
- Gemini CLI has issues with local image uploads and prompt interpretation.
Intro
The video introduces Google's Gemini CLI as a competitor to Cloud Code, aiming to compare their performance in a real-world production application. Gemini CLI offers a free tier with usage limits, which already makes it cheaper than Cloud Code. The video references an example where Gemini CLI outperformed Cloud Code in a stock tracker app, especially as the project grew in complexity. The presenter suggests that Cloud Code might be using Gemini 2.5 Pro, which has a smaller context window compared to Claude.
Getting Started
To begin using Gemini CLI, the presenter outlines the initial steps, including running a command, selecting a color scheme, and signing in with a Google account. The presenter notes that Cloud Code uses Bootstrap CSS for full-stack Next.js projects, which is considered less modern than Tailwind CSS. The presenter suggests explicitly specifying Tailwind CSS when creating new projects with Cloud Code.
Real World Example
The presenter uses Tensor AI, an application for AI news, as a real-world example. The goal is to add swiping functionality for navigating between news articles and replace 11 Labs with Miniax speech02 Turbo for audio generation to reduce costs. Both Gemini CLI and Cloud Code will be used to implement these features, and their performance will be compared.
The Prompt
The presenter duplicated the Tensor AI folder, creating separate folders for Gemini CLI and Cloud Code to run on the same files. The presenter shares the prompt used for both platforms and expresses the intention to compare their speed and quality of implementation.
Progress of Both
Gemini CLI quickly completes an initial change while Cloud Code is still navigating the directory. Gemini CLI switches from Gemini 2.5 Pro to Gemini 2.5 Flash due to rate limiting. Meanwhile, Cloud Code continues to navigate the codebase.
Claude Code’s Result
Cloud Code finishes its task, while Gemini CLI appears stuck after switching to the Flash model. The presenter tests Cloud Code's solution on an iOS emulator but encounters an issue with the pan gesture handler. The presenter provides Cloud Code with a screenshot of the error to fix. After the update, the swiping functionality works well, with a tracker indicating the current article. The presenter then tests the audio generation feature.
Review of Claude's Result
Cloud Code successfully implements the swiping feature with a minor correction. However, the initial audio generation fails due to an incorrect version specification in the replicate configuration. After fixing the version, the audio generation works perfectly. Overall, Cloud Code performs well but requires some debugging. Gemini CLI remains stuck after switching to the Flash model.
Gemini CLI’s Progress
After killing the Gemini CLI process, it's noted that it used a significant number of input tokens but few output tokens. The presenter compares this to Cloud Code's token usage, which appears more efficient. The presenter attempts to rerun Gemini CLI, trying to avoid the switch to the Flash model.
Upgrading Gemini CLI's Plan
Gemini CLI encounters the same issue, getting stuck after switching to the Flash model. The presenter decides to upgrade to the standard tier, which requires using Gemini code assist agent mode. Despite claims of high usage limits, the presenter finds it necessary to upgrade due to the free tier's limitations. The presenter successfully signs up for a free trial of the paid subscription.
Gemini CLI’s Progress (again)
After upgrading, the presenter reruns Gemini CLI with the same instructions. The process is faster, possibly due to caching. However, the resulting code contains a syntax error. The presenter attempts to provide a screenshot of the error to Gemini CLI, but it fails to recognize the local file. The presenter copies the failing code and pastes it into Gemini CLI for correction. The corrected code introduces new issues, including adding unnecessary code and misunderstanding the instructions. The presenter reverts changes and tries again, encountering the same issue as Cloud Code with the missing roots view. After fixing this, the swiping functionality works but is laggier and has direction issues compared to Cloud Code's implementation. The presenter then checks the audio generation.
Token Usage
The presenter checks Gemini CLI's token usage, estimating it to be around 700,000 tokens, after removing some earlier usage. The presenter notes that the token measurement might differ between Gemini CLI and Cloud Code, as the figures seem low.
Gemini CLI Commands
The presenter describes Gemini CLI commands, including the ability to edit code in an external editor like Cursor, VS Code, or Vim. Other commands include chat save to save and resume chats, chat list to view available chats, compress docs, memory show to view user preferences, and theme to change the theme. The presenter also lists available tools like text search, file writing, web search, and Google search.
My Conclusion
The presenter concludes that while Gemini CLI has a nice free tier, its performance degrades significantly when it switches to the Flash model due to usage limits. Gemini CLI requires more guidance and produces less satisfactory solutions compared to Cloud Code in the real-world example. The presenter plans to continue using Cloud Code unless Gemini CLI receives significant updates to improve its capabilities. The presenter also hopes for improvements to the billing process, which is currently confusing.
