15,000 stars a day! Its code generation outperforms Claude. Even Cursor is getting nervous? Google Gemini CLI is on a rampage!

Today, Google officially launched Gemini CLI, a version of its AI assistant for the terminal environment. The highlight of this tool lies in its extremely generous free usage quota: it supports 60 model calls per minute and 1,000 calls per day.

Is the Free Quota "Cheating" and Driving Claude Crazy?

Following Claude Code in February and OpenAI Codex CLI in April, Google also launched Gemini CLI in June. So far, the three major AI labs have all released their "terminal intelligent agents" tools - these CLI tools can read and modify files and execute commands on behalf of users in the terminal.

Many people probably thought that such terminal tools would always be a niche area, but now it seems that its potential has been underestimated. Many developers spend hundreds or even thousands of dollars on Claude Code every month, indicating that this "niche market" is actually much larger and more important than expected.

Compared with Claude Code, Google is quite "generous" in terms of price:

Log in with a personal Google account, and you can get a free license for Gemini Code Assist. This will unlock the Gemini 2.5 Pro model and a million-token context window. During this preview period, Google offers the most lenient call quota in the industry: up to 60 requests per minute and up to 1,000 requests per day, all for free.

Currently, the community's focus on Gemini CLI is mainly on its huge free quota. Commenters said bluntly, "It's too exaggerated. This will put great pressure on Anthropic." "Indeed - if these call limits are real and free, then there is finally real competition."

Different from Claude Code, Gemini CLI, like OpenAI Codex CLI, is open source (using the Apache 2.0 license). And Gemini CLI gained 15.1k stars in less than a day.

The functions provided by Gemini CLI include code writing, problem debugging, project management, document query, and code explanation. It is also connected to the MCP (Model Context Protocol) server and has Agentic AI capabilities.

The advantage of CLI tools is that they can be used with any editor or IDE, not limited to tools supported by specific plugins; at the same time, they also support concurrent operation of multiple instances. Some developers also believe that command-line interaction is more efficient.

Gemini CLI supports Mac, Linux (including ChromeOS), and Windows platforms. Different from Claude Code or Codex, it is natively implemented on Windows without relying on the Windows Subsystem for Linux (WSL). Developers can customize the context and parameters by adding a text file named gemini.md in the project root directory. Google senior engineer Tayor Mullen said that when the CLI detects "details worth saving for a long time", it will also automatically write them into this file.

Because it is open source as a whole, we can also see its system prompt.

As always, this system prompt not only defines the tool's behavior but also serves as a very concise and accurate user manual. For example, its principle for code comments is:

Comments: Please add comments carefully and focus on explaining "why" you are doing something, especially when dealing with complex logic, rather than explaining "what" you have done. Only add high-value comments when they can truly improve readability or when the user specifically requests them. Do not modify comments that are not related to your changes, and never communicate with the user or explain your changes through comments.

The technology stack used by the system by default is also quite valuable for reference.

If the user does not specify a technology preference, the default suggestions are as follows:

Web Front-End: React (JavaScript/TypeScript) + Bootstrap CSS, combined with the Material Design specification;

Backend API: Node.js + Express.js or Python + FastAPI;

Full-Stack Application: Next.js (React/Node.js) + Bootstrap + Material Design, or Python (Django/Flask) + React/Vue.js front-end;

Command-Line Tool (CLI): Python or Go;

Mobile App: Compose Multiplatform (Kotlin) or Flutter (Dart) for cross-platform development; also supports native development with Jetpack Compose (Android) or SwiftUI (iOS);

3D Game: HTML/CSS/JavaScript + Three.js;

2D Game: HTML/CSS/JavaScript.

What Makes Gemini CLI Better Than Specialized Tools?

Gemini CLI is connected to Google's most advanced coding and reasoning model, Gemini 2.5 Pro, which has the capabilities of code understanding, file operation, command execution, and dynamic troubleshooting, comprehensively improving the command-line usage experience. Users can write code, debug problems, and optimize workflows through natural language.

Moreover, Gemini CLI can also call other Google services based on MCP to generate images or videos in the terminal, realizing an integrated experience from coding to creation.

The integrated demonstration at the press conference and the underlying technical concept actually reveal the key positioning of Gemini CLI - it is not just a code-writing tool but a "terminal intelligent agent" that connects multimodal intelligence and bridges coding and creation. This concept is also reflected in the selection of the underlying Gemini 2.5 Pro model.

It is worth noting that Gemini 2.5 Pro itself is not a "code-specific model" specially trained for code tasks but a general model that covers a wider range of capabilities. So, why didn't Google create a customized code model for Gemini CLI like some of its competitors? In an in-depth interview, product manager Connie Fan and research manager Danny Tarlow shared their thoughts.

Connie Fan said that although specific tasks can indeed benefit from specialized models, for example, the code completion model trained by Cursor "performs very well in very narrow use cases", she emphasized that for most real-world development tasks, coding requirements are often not limited to the code itself.

For example, she mentioned a classic "vibe check" request: "Help me create a Taylor Swift song ranking app" - this requires the model not only to understand code but also to understand the context, have common sense, and even have some UI/UX aesthetics. "These common-sense requirements actually represent the real programming needs of users," she concluded. "Most general tasks cannot benefit from a pure code model."

Danny Tarlow further pointed out: "What exactly does a code-specific model mean? 'Code' is no longer just the code itself; it encompasses various aspects of the software development process and involves multiple information sources, some specific to code and some not. If we only strengthen code capabilities while weakening other capabilities, it will actually limit the model's performance. We prefer the coordinated development of general models, seeking the integration and balance of different capabilities to create a 'jack-of-all-trades' model, which is a better development path."

In fact, more and more developers have begun to notice the significant improvement of Gemini 2.5 Pro in code generation and understanding.

In an environment where Claude has long dominated the discussion, many people have started to wonder, "Is Gemini going to make a comeback?" and "Is Google going to turn the tables?" On social media, some users have shared their development experiences: in a project with about 500,000 lines of code, the quality of the code generated by Claude Code "is far inferior to that of CLI + Gemini 2.5 Pro". To achieve the same effect, developers need to keep an eye on Claude throughout the process.

Some developers also said that the completion rate of programming with Gemini 2.5 Pro on Trae has significantly exceeded that of Claude 3.7. Others said that Gemini fixes bugs much faster than Claude Code.

For a 3D rendering problem, Claude Code spent two hours trying to fix the bugs but still couldn't solve it. When I asked Gemini 2.5 Pro, it solved the problem in just 5 minutes. It seems that for those seemingly difficult problems, we can first let Gemini write a detailed instruction manual and then let Claude Code implement it according to the manual. This combined usage might be quite good.

These feedbacks indicate that in the past year, Google's improvement in code capabilities has entered a "qualitative change period". Connie Fan, the product manager of Gemini coding products, said that the key to the good reputation of Gemini 2.5 Pro lies in the systematic evolution of two aspects: "data" and "methodology".

From a data perspective, "code repository context" has become very crucial. Now the goal of the model is not just simple code completion but to understand and modify large codebases distributed across multiple files and involving multiple modules: "The model should be able to make complex changes that you would make if you spent an hour sitting down and working on a familiar codebase."

To enable the model to truly understand these contexts, Google has also begun to systematically tap into its internal engineer resources.

This includes people like Jeff Dean. "They represent a new standard of 'capability level'," Connie said bluntly. "We have hundreds of thousands of the smartest, most outstanding, and sometimes most opinionated engineers in the world." With their feedback on different languages, technology stacks, and experience levels, Google can cover a wider range of use cases and improve the model's ability to match the "refined tastes of professional developers". "We didn't really make good use of this in the past, but now we're starting to do it, and the results are amazing."

From a research perspective, the Gemini team did not put all their eggs in the "single large context window" basket but explored a dual-track approach: on the one hand, continuously expanding the context length; on the other hand, developing an agentic programming model with autonomous search and reasoning capabilities. The model interacts with users in the form of an agent and even has a certain degree of autonomy.

As research manager Danny Tarlow described:

"If we throw you (a human developer) into a large codebase, what would you do? You would use code search, look at the file structure, click around, read some code, and then search for other information. The agentic model mimics this approach."

What's even more interesting is that when the model successfully handles a million-line codebase, users will experience a "magical" surprise. This is exactly the direction that Gemini 2.5 Pro is aiming for: not only to mimic the way humans work but also to break through human experience limitations and develop new "non-human" solutions.

For this reason, more and more developers have begun to pay attention to Google's comeback in this AI programming competition. Whether Gemini can "turn the tables" remains to be seen, but the signal of "catching up" is becoming clearer and clearer.

Reference Link:

https://www.youtube.com/watch?v=jwbG_m-X-gE

This article is from the WeChat official account "AI Frontline" (ID: ai-front), written by Tina and published by 36Kr with permission.