Recently, I’ve added a new partner to my development work—OpenAI’s Codex. Previously, my coding companion in the terminal was Claude Code, which I used for over half a year. But honestly, AI programming tools are not like relationships—having feelings alone is not enough; they need to deliver results.

The reason for this change is simple. Last month, I had a refactoring task for an old project that took several weeks to schedule. Every time I opened the terminal, started Claude Code, described the problem, and waited for it to analyze step by step… then suddenly—“You have reached the usage limit for the current session.” The context was lost, the reasoning chain was broken, and that lengthy troubleshooting process felt like it never happened. You could only start a new session, restate the problem from scratch, and watch it fumble through the project structure like an intern with amnesia.

One day before leaving work, I saw a post from Sam Altman stating that enterprise users switching to Codex within 30 days would receive two months free. I thought, why not give it a try? After a week of using it, I want to share my thoughts on these two AI programming tools and how they differ.

Don’t Rush to Ask “Which is Better?” Understand They Are Completely Different

Many friends ask me, “Which is better, Codex or Claude Code?” This question is inherently narrow. By 2026, these two tools are no longer simply about “which is better”—they represent two completely different execution forms. Choosing one depends on what type of developer you are and what kind of projects you work on.

To put it simply, Claude Code is like having a senior architect by your side, discussing and coding together, confirming each step with you before proceeding. Codex is more like having a versatile full-stack engineer; you throw a task at it, and it sits down to complete it independently, then submits a pull request for your review.

This analogy is not just made up. Claude Code runs in the terminal, reads your local files, executes your local commands, and directly interacts with your Git repository—throughout the process, you can see what it’s doing. In contrast, Codex operates in a cloud-based sandbox environment; your code is cloned into an isolated container where it writes code, runs tests, and fixes bugs, then pushes the results back to you.

In summary: Claude Code follows a “human-in-the-loop” approach, while Codex adopts a “task delegation” model. Their design philosophies diverge from the start.

Advantages of Claude Code: Thoughtful but Needs Your Presence

So why have I used Claude Code for over half a year? Because it truly has some impressive capabilities.

First, the code quality is genuinely high. In the SWE-bench benchmark, which measures real programming ability, Claude Code once achieved a top score of 80.9%, while Codex was around 49%. Although Codex later improved to the 69%-80% range, Claude’s precision and reliability when handling complex codebases and cross-file refactoring is reminiscent of a seasoned expert. Once, I asked it to help refactor a payment module with over 2000 lines of code, and it not only modified the code but also added unit tests, wrote migration documentation, and highlighted several potential performance issues. This global oversight is Claude Code’s standout skill.

Second, it has strong memory and context capabilities. Claude Code has a unique feature called “CLAUDE.md”—you place a markdown file in the project root directory that outlines your team’s coding standards, architectural decisions, and common commands, and it reads this file every time it starts. It also employs an automatic context compression technique that summarizes previous content when the dialogue reaches 50% capacity, achieving a near “infinite conversation” effect. For teams that need to maintain large codebases over time, this is not just a tool but a long-term partner that remembers project history.

Third, the ecosystem integration is quite powerful. Claude Code supports the MCP protocol, allowing direct connections to Jira, Slack, Google Drive, reading requirement documents, updating tickets, and sending notifications seamlessly. It can also run directly in GitHub Actions and GitLab CI/CD pipelines, automating code reviews and issue categorization effortlessly.

But Why Did Claude Code Make Me Switch?

Given Claude Code’s strengths, why did I still switch? This brings us to its “three sins.”

The first issue is the quota. This is perhaps the most common complaint I’ve heard. Pro users have reported using up to 60% of their 5-hour session quota in just 3 minutes, and Max 20x users (paying $200 per month) saw their usage jump from 21% to 100% after a single prompt. Worse still, Anthropic employees later explained that during Pacific Time business hours from 5 AM to 11 AM, quotas would deplete faster. In other words, you pay the same amount, but if you work during peak hours, the service will run out more quickly. For developers who rely on programming for their livelihood, this is not just an “experience issue”; it’s a “livelihood issue.”

The second issue is more critical—diminished intelligence. In April 2026, AMD’s AI director Stella Laurenzo conducted a quantitative analysis. She analyzed 6,852 sessions and 235,000 tool calls, concluding that Claude Code’s depth of thought dropped by 67%, and the file reading rate before code modifications fell by 70%, with instances of poor behavior skyrocketing by 173%. Even more bizarre, some users found that Opus 4.7 could incorrectly answer basic tests like “how many r’s are in strawberry” and admitted to being “a bit lazy” for not performing cross-validation. A coding agent lost its qualification to be entrusted with terminal tasks. Although Anthropic later issued an apology acknowledging that three bugs compounded the issue, trust, once lost, is hard to regain.

The third issue is the double-edged sword of local execution. Claude Code runs on your local machine, directly interacting with your file system—this is both an advantage and a risk. It can see everything in your terminal. In contrast, Codex runs in a cloud sandbox, executing tasks in an isolated container that is destroyed after use. For developers handling sensitive code, this sense of “isolation” can be more reassuring.

Advantages of Codex: Fast, Affordable, and Less Supervision Required

On the first day I switched to Codex, my most immediate impression was—smooth.

First, the startup and response speed is noticeably faster. The Codex CLI is written in Rust, resulting in a small binary file and quick execution speed. For the same instruction to “analyze the structure of this project,” Codex produces results almost instantly. When handling simple tasks, Codex consumes about one-third the tokens of Claude Code, leading to significant cost savings.

Second, the fully open-source ecosystem is a big plus. The Codex CLI is open-sourced under the Apache-2.0 license, making it fully auditable, which is a real benefit for corporate compliance teams. Moreover, Codex has already integrated with mainstream IDEs like VS Code, Cursor, and Windsurf. If you’re already using ChatGPT Plus or Pro, you can log in with your existing account without needing to register an additional Anthropic account. Not to mention it has recently made its way to mobile—allowing you to monitor Codex’s task progress, approve commands, and switch models right from your phone, which is a lifesaver when traveling or needing a quick check.

Third, the “task delegation” model creates a flow state experience. Claude Code requires confirmation for every step—reading files, modifying files, running commands. While this is necessary for high-risk operations, it becomes a distraction when you have to handle multiple small tasks throughout the day. Codex delegates tasks to the cloud, allowing you to focus on other things while it notifies you when it’s done. For someone like me, who needs to juggle three or four projects, this asynchronous work style better matches my rhythm.

Of course, Codex has its own issues. When it comes to multi-file cross-module refactoring, it sometimes focuses too narrowly, modifying one file while missing calls in another. For complex architectural decisions, Claude Code remains the more reliable choice.

Realizations: It’s Not About Choosing Sides

After a week of use, my biggest realization is—good developers should not rely on just one tool.

NVIDIA CEO Jensen Huang mentioned at GTC 2026 that NVIDIA internally uses AI programming tools 100% of the time, and they use Claude Code, Codex, and Cursor together. I believe this is the right approach.

My current practice is: use Codex for small tasks, daily development, and quick prototypes; use Claude Code for complex refactoring, architectural design, and troubleshooting strange bugs; and use Cursor for daily editing and completion. Tools should not become a belief; they are merely means to help you get the work done well.

Finally, I want to emphasize that regardless of which one you use, the most important thing is—you are still the one making decisions. AI can help you write code, run tests, and fix bugs, but it will not replace your thinking about “why this feature should be implemented,” nor will it judge whether “this architectural decision will become technical debt in five years,” and it certainly won’t take on the responsibility for deploying the code. The industry is changing, tools are evolving, but the one responsible for your code will always be you.