Cursor falls from the top 3, and OpenAI's Codex rises to the top 3 in terms of market share. What will happen to it? ～Summary of AI Coding trends in September 2025 based on data～
AI Coding Agent September trends
We have been operating a site called AI Coding.Info since July 2025.
This is a site that observes usage trends related to AI Coding Agents such as Claude Code, Gemini, or Codex from a fixed point from information in Github repositories. To determine the use of AI Coding Agent, we conduct daily surveys under the following conditions.
- Daily inspection of 9,000 public repositories on Github
- TOP 300 Github stars by programming language
- Researched from 30 types of programming languages
- Targets 16 types of AI Coding Agents
- Only when the rule files used by various AI Coding Agents are in the Github repository, it is determined that AI Coding Agent is being used.
Past trends
https://qiita.com/kotauchisunsun/items/092784402a36d5705853
https://qiita.com/kotauchisunsun/items/a1e06dd590f945ae09ef
AI Coding Agent usage rate is 3.6%
**The AI Coding Agent repository usage rate at the end of September was 3.6%, an increase of 0.5 points from 3.1% last time. ** Last time, the increase rate from July to August was 0.2 points. Although the rate of increase is increasing, it has not changed significantly. That's the impression.
https://ai-coding.info/ja/agents
Share of AI Coding Agent by product
The share by product is as follows.
| Ranking | Product name | Share rate |
|---|---|---|
| 1st place | Claude Code | 31.8% |
| 2nd place | Copilot Agent | 26.7% |
| 3rd place | Codex CLI | 16.1% |
| 4th place | Cursor | 14.8% |
| 5th place | Gemini CLI | 6.3% |
https://ai-coding.info/ja/agents
**The market share of Claude Code and Copilot Agent continues to be high. That fact remains the same, but the overall market share is decreasing by about 1%. That's what it feels like. Cursor, which was in 3rd place last month, has moved up to 4th place, and Codex CLI, which was in 4th place, has moved up to 3rd place. ** Last month, I wrote an article with the following content.
Also, although Cursor has a higher number of repositories than Codex CLI and Gemini CLI, the growth rate is low, so it is possible that Cursor may drop in the rankings. If we assume a fixed growth rate, Codex CLI may surpass Cursor in the number of repositories one month later, in September.
In this way, **Cursor, which was originally in 3rd place in market share, will be overtaken by Codex CLI in 4th place in September. That was my prediction. In fact, on September 25th, the rankings were reversed. **
https://x.com/kotauchisunsun/status/1971170568461508839
I felt that the release of GPT-5-Codex on September 15 was a trigger.
https://x.com/OpenAI/status/1967636903165038708
Since this release, within my observation range, I've seen some comments saying, ``I tried Codex.'' It's a clever idea. There was also this opinion. In fact, it seems to be getting good scores on benchmarks.
Quote: https://artificialanalysis.ai/?intelligence-tab=coding#artificial-analysis-coding-index
According to https://artificialanalysis.ai, currently the best model for coding is Grok 4, followed by GPT-5 Codex(high) and GPT-5(hight). In the graph above, since it is a composite index, the first place is Grok 4, but GPT-5 Codex (high) may be the first place in individual coding benchmarks. Therefore, the performance is slightly better than Claude 4.5 Sonnet. What it feels like may actually be true. However, I haven't heard that they have replaced the Claude Code they use internally with Codex, so it's still a bit more of a wait-and-see approach, or it's not enough to replace Claude Code. Replacing it is costly. Therefore, it does not exceed that. It may be an evaluation like that. Previously, I had written a [review article of Codex CLI] (https://qiita.com/kotauchisunsun/items/fa4a3167c2480984216b), but when I touched it for the first time in a while, the behavior had changed, and the contents of the previous verification no longer worked. Therefore, it is necessary to periodically check the operation.
AI Coding Agent usage status by programming language
**The programming language in which AI Coding Agent is used most is "Typescript," followed by "Rust" and "Python." ** As you can see from the two-month graph, the ranking has not changed from TypeScript to JavaScript. Therefore, there has been a clear distinction between programming languages that are suitable for AI Coding and languages that are not suitable for AI Coding. You might say that.
https://ai-coding.info/ja/agents
The graph below compares the usage rate of AI Coding Agent in the TypeScript repository and the usage rate of AI Coding Agent in the Rust repository as of October 1st.
TypeScript and Rust are ranked first and second in terms of number of repositories used, but the reality is slightly different: TypeScript is used in 23.7% of repositories, while Rust is used in only 11.7%. As you can see, there are considerable differences in usage between programming languages, with a gap of more than 10% between the first and second place. Going a little deeper into the discussion, when it comes to TypeScript, it has been confirmed that one in five repositories uses AI Coding Agent. Regarding TypeScript, the use of AI Coding Agents has become quite common not only from the experience but also from the data. I think it's fair to say that. **
https://ai-coding.info/ja/languages/typescript
https://ai-coding.info/ja/languages/rust
Monthly trends in the number of repositories used by AI Coding Agent
The number of repositories as of September 1, 2025 is 374, and the number of repositories as of October 1, 2025 is 446, an increase of 72. Although there are some fluctuations depending on the day, the number of repositories using AI Coding Agent is steadily increasing, as can be seen from the date and time trends. As I discussed in last month's article, the black part at the bottom is the Cursor, and the green part is the Codex CLI. This figure also shows that the number of repositories for Cursor is constant, while the number of repositories for Codex CLI is increasing.
https://ai-coding.info/ja/agents
OpenAI and Windsurf/Cursor. And AGENTS.md
OpenAI released the Codex on April 17, 2025. There were several timely articles around that time. One is OpenAI's acquisition of Windsurf.
https://www.publickey1.jp/blog/25/openaiaiwindsurf.html
On May 7, 2025, OpenAI acquired Windsurf for the equivalent of $3 billion (450 billion yen). There was a story. However, looking at overseas news sites, OpenAI wanted to acquire Cursor, not Windsurf. There are also rumors. This is the news for April 17, 2025.
Therefore, OpenAI was considering entering the field of AI Coding, and wanted the reach, developers, and customers of early movers such as Windsurf and Cursor. You can make an analogy. However, it was actually Windsurf that we were able to acquire, not Cursor. On the other hand, the flow AGENTS.md was born on August 20, 2025.
https://gihyo.jp/article/2025/08/agents-md-site
In AI Coding Agent, each tool has a document file for instructing AI with information in the repository. For GEMINI, it is GEMINI.md, for ClaudeCode, it is CLAUDE.md, etc. The movement to unify this is AGENTS.md, and OpenAI is leading the initiative. As of October 2, 2025, the following 18 types of AI Coding Agents are listed on the official website.
Looking at this, it seems that AGENTS.md is compatible with a fairly wide range of issues, but there are various points that are puzzling.
- No Windsurf
- Pure competitors such as Gemini CLI, Github Copilot, and Cursor are also supported.
As for Windsurf, its usage is quite low within the observation range of AI Coding.Info. Windsurf ranks 7th in the overall ranking, and ranks around 3 out of 9,000 repositories adopted.
https://ai-coding.info/ja/agents/windsurf
When you think about it that way, it doesn't feel like it has the same momentum as it did for a while. However, considering that OpenAI has acquired the development team itself, it would not be surprising if something like OpenAI's Windsurf, a VSCode-based IDE, comes out from OpenAI. In fact, Codex is also published as an Extension for VSCode, so that may be the direction. Considering this, Windsurf will be released with a new name, such as Codex IDE. In a sense, I think there is a possibility that the conventional methods of the field will emerge.
Regarding 2., what kind of industry structure is it? I think that is the focus. ClaudeCode is probably being used a lot now. There is a sense that this is a possibility, and I think each AI Coding Agent development company is developing countermeasures. That's what I think. On the other hand, although Github is lagging behind Anthropic, it cooperates with other AI Coding Agents to compete against ClaudeCode due to its strong reach. Personally, I thought it was difficult to imagine. Similarly, I thought it would be difficult for Google and others to participate in such a siege against ClaudeCode due to their strength. So, when I dug deeper into this AGENTS.md area, I found something interesting. Take a look at the following article that introduces Github Copilot.
https://aadojo.alterbooth.com/entry/2025/09/06/131024
Priority with copilot-instructions.md What I'm curious about here is which of Copilot's traditional custom instruction files, copilot-instructions.md and AGENTS.md, takes precedence. The Japanese version of the document below states the following:
If you choose not to include a Copilot instruction file in your repository, Copilot will fall back to your existing custom instructions, such as CLAUDE.md, AGENTS.md, and GEMINI.md.
Recent Github Copilot documentation no longer mentions fallback, but I guess that's probably the case. I think so. Github Copilot also seems to be gaining a lot of shares, and the documentation says,
Copilot Coding Agent supports different types of custom instruction files.
/.github/copilot-instructions.md /.github/instructions/**/*.instructions.md **/AGENTS.md /CLAUDE.md /GEMINI.md
You can see the attitude of going to see what other companies' rule files are. Also, regarding GEMINI CLI, the following description can be found in AGENTS.md.
What this means is that **Gemini CLI only sees GEMINI.md by default. Once set, it will read AGENTS.md. **
That being said, I could sense that things were going on quite a bit under the surface. If the rumors are true, Cursor is in a perfect position, but there is also a document that describes how to deal with AGENTS.md.
Reference: https://docs.cursor.com/ja/context/rules
My impression is that AGENTS.md can be used as a simple replacement for .cursor/rules. I feel a bit stuck in this area. Therefore, while AGENTS.md appears to be a concerted effort in the industry to take on ClaudeCode, it is aiming for the top spot in the industry as if it were trying to outwit a living horse. You can even get a glimpse of that. This is common sense, and if this area is standardized by AGENTS.md, switching between AI Coding Agents will go smoothly. There is a problem. This means that the switching cost for users will be zero, increasing the possibility that they will be able to discard the AI Coding Agent they are currently using. In other words, this means that once a good AI Coding Agent is released, users will quickly move on to other products, which means that competition will intensify from the perspective of AI Coding Agent developers. I don't know if this is a good idea, but if that happens, companies with huge capital will either compromise on the quality of the software, or they will lock in with LLM performance like OpenAI. Either way, I can see a future in which other AI Coding Agents will disappear as we are drawn into a battle between Big Techs. If you think about it that way, AGENTS.md can be seen as a way to absorb users of other AI Coding Agents through OpenAI, and its share is actually increasing, but Github and Google are also taking solid defensive measures against this. That is my impression (Note: AI Coding.Info calculates the repository used by AGENTS.md as the share of Codex CLI). On the other hand, there seems to be some AI Coding Agents that are integrated into AGENTS.md, but I think this is strategically okay. There is a part of me that thinks so.
thoughts
Regarding the trends in September, there were some parts that diverged from our intuition. this time,
The AI Coding Agent repository utilization rate at the end of September was 3.6%, an increase of 0.5 points from 3.1% last time. Last time, the increase rate from July to August was 0.2 points. Although the rate of increase is increasing, it has not changed significantly. That's the impression.
I wrote a sentence called . Admittedly, it is numerically small. However, the fact that this growth rate is increasing by 0.3% seems to strike a chord in my heart, and I have a gut feeling that I should not ignore it. I feel that the coding speed is increasing rapidly with AI Coding Agent, but based on the data from the past few months, it is not that noticeable. That was my opinion. However, at this point, we are starting to feel that the popularity is accelerating and expanding exponentially, even from the data. In fact, we have confirmed that TypeScript is becoming more popular at a fairly rapid rate. Is this area increasing at an accelerating pace? That means things will become clear within the next month or so. I think so. However, I think there is another epoch-making point of evolution. I wrote an article like this before.
https://qiita.com/kotauchisunsun/items/c8b6484c98557da44893
This is my thesis, and the gist is as follows.
"One of the factors behind the popularity of modern programming languages is that they have a solid package manager. To achieve this, it is important that the reusability and independence of modules in a programming language are maintained. Programming languages that meet these requirements are still popular today because their reusability has accelerated with the spread of the Internet."
I feel that programming languages are at an inflection point with AI Coding Agent. C/C++ has not been widely used in AI coding for some time. But Rust is expensive. Considering this, I think that low-level languages such as embedded languages will also undergo a revolution with the introduction of AI Coding Agent. In addition, the adoption rate of AI Coding in Java is very low. I think Go, Rust, and Java are the mainstream on the large-scale server side these days, but this may change again. In a sense, AI Coding Agent is an ``easy to write'' programming language and its surrounding mechanisms. Elemental technologies such as the combination of the Internet and package managers may emerge in the future. This could be MCP for one thing, or it could be a spec-based approach like Kiro or Spec Kit, which have recently become popular.