Exa is a search tool designed for AI, providing an API that allows models to perform web searches and summarize content from web pages.
Exa's service is priced at $5 per 1,000 requests, which is significantly cheaper compared to Google's search grounding for models at $35 per 1,000 requests.
Gemini CLI brings the power of Gemini directly into developers' terminals, offering "unmatched access for individuals."
It's an open-source AI agent designed for a wide range of tasks, including coding, content generation, problem-solving, deep research, and task management.
The CLI integrates with Google's AI coding assistant, Gemini Code Assist, supporting both VS Code and the Gemini CLI.
Google is acknowledging existing development standards like VS Code and offering an OpenAI-compatible API, a shift from its previous approach.
Google leverages its extensive resources, including its compute platform, infrastructure, and financial capital, to provide very generous usage limits.
The free tier of Gemini CLI allows for 1,000 requests per day, even with a 1 million token context window.
Based on an estimated 200,000 input tokens and 8,000 output tokens per request, 1,000 daily requests could cost Google approximately $500 for input tokens and $120 for output tokens, totaling over $620 per day.
This aggressive free tier is seen as a "hostile" move towards competition, as Google is effectively absorbing significant costs.
A conceptual "battle royale" experiment pitted various coding agents (Open Code, Claude Code, Codeex, Gemini) against each other in a simulated task to "kill all other processes."
In the first round, Open Code won quickly by immediately identifying and terminating other processes.
The experiment highlighted the distinct planning and execution strategies of different AI agents.
Gemini CLI is a TypeScript project, built by a "modern team" that embraces VS Code and existing development standards.
Its architecture separates CLI and core packages, with the core handling Google authentication and other foundational elements.
The CLI package utilizes Ink, a React-like framework for building command-line interfaces, which allows for faster development but can sometimes impact performance compared to CLIs built with Go or Rust.
Competitors like Codeex are exploring Rust rewrites to enhance performance and eliminate Node.js dependencies, though this might slow down iteration.
The OpenAI version of Codeex has already incorporated support for non-OpenAI models, including Claude and Gemini.
The presenter tested Gemini CLI with a task to modify a UI to display two image placeholders for Gemini's image generation (which produces two images), similar to how OpenAI's single-image generation is handled.
Gemini CLI initially produced a vertical stacking change, requiring a follow-up prompt to achieve the desired horizontal layout.
This task consumed approximately 400,000 input tokens and took about 1 minute 30 seconds of API time, estimated to cost around $8 if charged.
Comparing Gemini CLI with Claude Code & Open Code 23:36
Claude Code: Took significantly longer (nearly 5 minutes for the initial step, almost double Gemini's total time) and required more manual guidance (e.g., specifying to add image count as a field). While using fewer tokens overall, it had high cache read/write activity and cost approximately $1.
Open Code (SST/open-code): Was tested last and completed the task correctly on the first attempt without additional prompting or interruptions.
In the comparison, Open Code was the fastest and cheapest, costing only 5 cents for the task, significantly less than Gemini CLI and Claude Code, which cost over a dollar each.
This suggests Open Code might be more efficient in its token usage or context handling.
Gemini CLI's system prompts are open source and serve as clear documentation of its capabilities and preferred behaviors.
The prompts guide the model on best practices for adding code comments (sparingly, focusing on 'why'), and suggest preferred technologies like React, TypeScript, and Bootstrap CSS for frontends, Node/Express or Python/FastAPI for backends, and Next.js for full-stack.
Building the CLI allows Google engineers to directly experience Gemini's "quirky behaviors," particularly regarding tool calls, and serves as a testbed for improving model reliability.
Google is actively working to fix issues with Gemini 2.5 Pro's tool calling, which is noted as being less reliable than older 2.0 models.
The free tier's data usage policy for training on user code is not explicitly stated, implying it might be used for improvements unless a paid API key is provided.
Model companies building CLIs (e.g., Gemini CLI, Codeex, Claude Code, Open Code) is crucial for "dogfooding" their models and gaining a deeper understanding of their characteristics.
This process forces internal teams to orchestrate model components, identify strengths and weaknesses, and become more effective users of their own models.
The creation of these CLIs directly contributes to model improvement, especially in areas like tool calling.
Gemini CLI is highly open and customizable, allowing users to define and use their own custom system prompts (e.g., via Gemini/system.md).
It also supports extensions, enabling global or workspace-specific configurations for Gemini's access and functionality.
Each major CLI project (OpenAI Codeex, Open Code, Gemini CLI) has contributed to advancing the state-of-the-art in different areas, such as open-source practices, user experience, and customization.
The transparency and active engagement from lead developers are positive indicators for the future development of these tools.