---
date: "2025-09-03T21:07:10.000Z"
title: "2025-09-03"
tags: ["codex","gpt-5"]
draft: false
---

I tried OpenAI's `codex` CLI coding agent again today for the first time in a _while_.

```sh
Update available! 0.1.2504301751 → 0.29.0.
```

I recently shipped a release of [Tomo](https://www.wvlen.llc/apps/tomo) but there was a minor bug with the horizontal scrollbar overlapping a codeblock with syntax highlighting.

I prompted `codex` to fix this and it seemed to localize the problem and suggest a working fix in one attempt.

> ensure that the horizontal scrollbar in styled codeblocks does not overlap the code itself. right now I am seeing that behavior

I was motivated to try this out because I've been hearing rumblings about `gpt-5` being good at coding.
It's currently the number 4 coding model on OpenRouter and climbing, relative to both Sonnet and Qwen3 Coder.

| Rank | Model                 | Company   | Parameters   | Performance |
| ---- | --------------------- | --------- | ------------ | ----------- |
| 1    | Grok Code Fast 1      | x-ai      | 844B tokens  | ↑ 3,231%    |
| 2    | Claude Sonnet 4       | anthropic | 339B tokens  | ↓ 9%        |
| 3    | Qwen3 Coder 480B A35B | qwen      | 112B tokens  | ↓ 33%       |
| 4    | GPT-5                 | openai    | 66.2B tokens | ↑ 21%       |
| 5    | DeepSeek V3.1         | deepseek  | 60.9B tokens | ↓ 56%       |

I believe Grok Code Fast 1 was launched in some kind of free capacity along with its release which may in part explain its high usage but we will see where that lands.