6 min read
|
Saved February 14, 2026
|
Copied!
Do you care about this?
The article details the author's process of reverse engineering the Codex CLI tool to access the newly released GPT-5-Codex-Mini model. By modifying the tool, they created a command to generate SVGs based on user prompts, specifically focusing on drawing a pelican riding a bicycle. The author shares insights into the coding challenges and the functionality of the modified tool.
If you do, here's more
OpenAI recently launched GPT-5-Codex-Mini, a smaller and more affordable variant of their GPT-5-Codex model. Currently, access is limited to the Codex CLI tool and a VS Code extension, with broader API access pending. In a creative approach, the author reverse-engineered the Codex CLI, which uses Rust, to enable direct prompting of the new model. This method exploited the open-source nature of Codex CLI, allowing for experimentation without violating terms.
The author detailed the process of modifying the Codex CLI to introduce a new command called βcodex prompt.β This command allows users to send prompts directly to the OpenAI API, bypassing the default behavior of the Codex model that typically assumes a full workspace is available. Initial attempts to generate an SVG of a pelican riding a bicycle faced challenges, as the model continued to reference workspace tools instead of directly responding to the prompt. After refining the command and adjusting the system prompts, the author successfully executed it, culminating in generating SVG images.
The final tests with both GPT-5 and GPT-5-Codex-Mini revealed varying results, with the mini version not meeting expectations for SVG creation. The code modifications and the iterative process exemplified an innovative use of existing tools to push the boundaries of what the new model can do, highlighting both the potential and limitations of GPT-5-Codex-Mini.
Questions about this article
No questions yet.