Introduction

On June 27, Google officially launched Gemini CLI, a version of its AI assistant designed for terminal environments. This tool stands out due to its generous free usage quota: it supports 60 model calls per minute and up to 1,000 calls per day.

Competitive Free Quota

Following the releases of Claude Code in February and OpenAI Codex CLI in April, Google introduced Gemini CLI in June. All three major AI labs have now released their own terminal agent tools, which can read and modify files and execute commands on behalf of users.

Many initially thought these terminal tools would remain niche, but the reality shows a much larger and more significant market. Some developers reportedly spend hundreds to thousands of dollars monthly on Claude Code, indicating a broader demand.

In contrast to Claude Code, Google’s pricing is notably generous:

Users can log in with a personal Google account to receive a free Gemini Code Assist license, unlocking the Gemini 2.5 Pro model and a million-token context window. During this preview, Google offers the industry’s most lenient calling quota: up to 60 requests per minute and 1,000 requests per day, all for free.

Community interest in Gemini CLI has focused on its substantial free quota, with commentators noting, “This is outrageous; it will put immense pressure on Anthropic.” If these calling limits are real and free, it finally introduces genuine competition.

Features of Gemini CLI

Unlike Claude Code, Gemini CLI, like OpenAI Codex CLI, is open-source (under the Apache 2.0 license). Within less than a day, it garnered 15.1k stars on GitHub.

Gemini CLI offers functionalities including code writing, debugging, project management, documentation queries, and code explanations. It also connects to the MCP (Model Context Protocol) server, providing 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; they also support concurrent multi-instance operation. Some developers believe command-line interaction is more efficient.

Gemini CLI supports Mac, Linux (including ChromeOS), and Windows platforms. Unlike Claude Code or Codex, it is natively implemented on Windows without relying on the Windows Subsystem for Linux (WSL). Developers can customize context and parameters by adding a text file named gemini.md in the project root. Google senior engineer Tayor Mullen stated that when the CLI detects “details worth preserving long-term,” it will automatically write them to this file.

Due to its open-source nature, we can also view its system prompt.

This system prompt not only defines the tool’s behavior but also serves as a concise and accurate usage document. For example, its principles for code comments are:

Comments: Please add comments cautiously, focusing on “why” something is done, especially when dealing with complex logic, rather than explaining “what” was done. Only add high-value comments when they genuinely enhance readability or when explicitly requested by the user. Do not modify comments unrelated to your changes, and never communicate or explain your changes to the user through comments.

The default technology stack used by the system is also quite informative:

• Web Frontend: React (JavaScript/TypeScript) + Bootstrap CSS, following Material Design guidelines.
• 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 frontend.
• Command-Line Tools (CLI): Python or Go.
• Mobile Apps: Compose Multiplatform (Kotlin) or Flutter (Dart) for cross-platform development; also supports Jetpack Compose (Android) or SwiftUI (iOS) native development.
• 3D Games: HTML/CSS/JavaScript + Three.js.
• 2D Games: HTML/CSS/JavaScript.

What Sets Gemini CLI Apart?

Gemini CLI leverages Google’s advanced coding and reasoning model, Gemini 2.5 Pro, which enhances code understanding, file manipulation, command execution, and dynamic fault diagnosis, significantly improving the command-line experience. Users can write code, debug issues, and optimize workflows using natural language.

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

The integrated demonstrations and underlying technical philosophy revealed during the launch highlight Gemini CLI’s key positioning—it is not just a coding tool but a “terminal agent” that connects multimodal intelligence, bridging coding and creation. This philosophy is also reflected in the choice of the Gemini 2.5 Pro model.

Notably, Gemini 2.5 Pro is not a specialized model trained specifically for coding tasks but a general model covering a broader range of capabilities. Why did Google not develop a customized coding model for Gemini CLI like some competitors? In a deep interview, product lead Connie Fan and research lead Danny Tarlow shared their thoughts.

Connie Fan emphasized that while specific tasks can benefit from specialized models, such as Cursor’s trained code completion model excelling in narrow use cases, she stressed that most real-world development tasks often extend beyond just the code itself.

For instance, she mentioned a classic “vibe check” request: “Help me create a Taylor Swift song ranking app”—this requires the model to understand not just code but also context, possess common sense, and even some UI/UX aesthetics. “These common sense aspects reflect real user programming needs,” she summarized, “Most general tasks do not benefit from a purely code model.”

Danny Tarlow further pointed out, “What does a code-specific model really mean? ‘Code’ is no longer just the code itself; it encompasses various stages in the software development process, involving multiple information sources, some specific to code and some not. If we only strengthen code capabilities while weakening others, it may limit the model’s performance. We prefer collaborative development on a general model, seeking integration and balance between different capabilities to create a ‘versatile’ model, which is the better path forward.”

In fact, more developers are beginning to notice the significant improvements Gemini 2.5 Pro has made in code generation and understanding.

Amid Claude’s long-standing dominance in discussions, many are now asking, “Is Gemini making a comeback?” and “Is Google turning the tide?” Users on social media have shared their development experiences: in a project with around 500,000 lines of code, the quality of code generated by Claude Code was “far inferior to cline + Gemini 2.5 Pro,” requiring developers to monitor Claude throughout.

Some developers also noted that the completion rate for programming using Gemini 2.5 Pro on Trae has significantly exceeded that of Claude 3.7. Others mentioned that Gemini fixes bugs noticeably faster than Claude Code.

A 3D rendering issue took Claude Code two hours to fix a bug without resolution, while Gemini 2.5 Pro solved it in just five minutes. For seemingly challenging problems, it might be beneficial to let Gemini draft a detailed specification first, then have Claude Code implement it—this combination could work quite well.

These feedbacks indicate that over the past year, Google’s advancements in coding capabilities have reached a “qualitative change” phase. Gemini coding product lead Connie Fan stated that the positive reception of Gemini 2.5 Pro hinges on two aspects of systematic evolution: “data” and “methodology.”

From a data perspective, “code repository context” has become crucial. The model’s goal is no longer just simple code completion but to understand and modify large codebases distributed across multiple files and involving various modules: “The model must be able to perform complex changes that you would take an hour to do if you sat down, familiar with the codebase context.”

To enable the model to truly understand these contexts, Google has begun systematically tapping into internal engineering resources.

This includes individuals like Jeff Dean, who represent a standard of “new levels of capability,” Connie stated, “We have tens of thousands of the smartest, most outstanding, and sometimes most opinionated engineers in the world.” With their feedback across different languages, tech stacks, and experience levels, Google can cover a broader range of use cases, enhancing the model’s alignment with the “refined tastes of professional developers.” “We actually haven’t utilized this well in the past, but now we are truly doing so, and the results are astonishing.”

From a research perspective, the Gemini team has not placed all its chips on the “single large context window” approach but is exploring dual paths: continuously expanding context length while developing agentic programming models with autonomous search and reasoning capabilities, where the model interacts with users in an agent-like manner, even possessing a degree of autonomy.

As research lead Danny Tarlow described:

“If we were to drop you (the human developer) into a large codebase, what would you do? You would use code search, look at file structures, click around, read some code, and search for other information. The agentic model mimics this approach.”

Interestingly, when the model successfully handles a million-line codebase, users experience a sense of “magical” surprise. This is precisely the direction Gemini 2.5 Pro is focused on: not just simulating human work methods but also breaking through human experience limits to develop “non-human” solutions.

Consequently, many developers are beginning to pay attention to Google’s resurgence in this AI programming competition. Whether Gemini will “turn the tide” remains uncertain, but the signals of “catching up” are becoming increasingly clear.