Switching from Cursor to Codex for Over Half a Month: My Real Experience

In April 2025, I began to distinctly feel that AI was gradually becoming commercially viable in the coding domain. Although it was still in the realm of auxiliary programming at the time, I self-funded a subscription to the then-popular Cursor for $20 a month. Soon after, at the end of April 2025, my company started providing enterprise subscriptions for employees. Cursor even proactively refunded my unused balance, which initially left me with a very good impression of Cursor.

I continued using it until April 2026.

Overall, the experience was stunning. Compared to "traditional coding," Cursor can be described as a new era.

Cursor's best feature was its close integration with the coding environment.

Tab completion, Cmd+K, Agent, and Plan Mode were all within the editor, making it smooth to modify functions, complete pages, and explain code. The human directs, and AI assists. Of course, one could also have AI write the code in Cursor first, then make manual corrections in their familiar IDE. Additionally, Cursor allowed users to choose their preferred mainstream models; it was a third-party IDE, not strongly tied to any single model.

At that point, Cursor still had quite a few issues. For example, there wasn't a directly usable Cursor plugin in Android Studio (AS). To synchronize code positions between Cursor and AS and avoid re-locating code after switching back to AS, I specifically installed the IDE Sync plugin. Another example was Cursor's CLI, which had many pitfalls, and its initial recognition of MCP and later SKILLS had bugs.

Of course, none of these were real problems. AI is developing too fast, and Cursor's version iterations were also very rapid. Through continuous improvements in each version, coupled with the models' own constantly enhancing capabilities, starting in the second half of 2025, with Claude's rapid rise, I distinctly felt that Cursor's understanding of code and its coding ability significantly improved.

What struck me most was that previously, for business details, I might have needed to ask developers (honestly, our project's business logic is quite complex). As Cursor iterated, I gradually found myself asking Cursor directly. Later, I wrote less and less code myself; Cursor could handle almost all scenarios.

On February 14, 2026, I published an article on my public account, entirely based on Cursor. In a short time, with zero code throughout, from design to implementation, I wrote a fully functional mini-program with decent visual effects, even without prior WeChat mini-program development experience.

At that moment, I had a feeling that the "Dharma-ending age" of programming was drawing nearer.

In April 2026, Cursor's pricing and enterprise strategy began to become complex. Cursor shifted from fixed licenses to usage-based billing, which was more expensive and unpredictable. Based on this, the company decided that the China region would adopt OpenAI Codex to replace Cursor.

As people joked on Maimai: "Who would have thought, before we even lost our jobs, our company's Cursor got 'fired' first."

Starting mid-April 2026, we officially embraced Codex.

Codex immediately felt very different from Cursor. Cursor previously felt more like an IDE, an IDE that coexisted with code, just with AI capabilities on the side. But Codex felt like an Agent capable of handling various work scenarios and tasks.

Of course, starting in April 2026, Cursor's new version also underwent significant changes, continuously transforming from an IDE form to an Agent form, sparking widespread discussion in the community. I personally believe this path and direction are correct. It's the general trend.

Codex's capabilities are truly strong. Coding, in its positioning, should only be one of its core scenarios. Codex's update speed is even more frequent; recently, there's been a new version almost every two days. Its interaction paradigm directly reflects its ambition: to be a super Agent, orchestrating capabilities across various scenarios.

After using it for over half a month, Codex is indeed easy to use and can be end-to-end.

It gives me a feeling somewhat similar to the OpenClaw Agent approach, but it makes "connections" much simpler and easier to use. It doesn't require users to do a lot of complex configurations; instead, it uses built-in "plugins." Users just need to simply enable the plugins they want to directly connect to external tools or applications.

It can also enable powerful features like "Computer use." The recently released Codex Chrome plugin further enhances its ability to automatically operate browsers, ultimately enabling full automation of many workflows. Even complex tasks can be handled directly by Codex.

Of course, compared to Cursor, currently, looking solely at coding ability, overall, I feel Cursor + Claude 4.7 is still stronger than Codex + GPT-5.5. However, the difference is not significant, and it's basically sufficient.

One more thing worth mentioning: I previously wrote about GPT Image 2, and its image capabilities are incredibly powerful, which directly gives Codex strong image capabilities in image-related scenarios.

Let's make a final summary: Cursor first made me strongly feel the power of AI programming, while Codex made me truly feel for the first time that AI is not just a programming assistant, but a true Agent. It's more powerful and easier to use than previous Agent tools like "Lobster" (龙虾). It's a true super Agent, and Codex is clearly closer to the next stage of AI tools.