Video tutorial · The Metaverse Guy
Testing GPT-5 Codex via Kilo Code: Faster, Cheaper Coding
I accessed GPT-5 through the open-source Kilo Code agent for VS Code and compared it against GLM-4.6, Claude Sonnet 4.5, and Kimi K2.
- Published
- 9 October 2025
- Length
- 6 min
- Views
- 1,958 views
In this video, I show how to use the new GPT-5 Codex model for coding tasks by pairing it with Kilo Code, an open-source AI agent for VS Code. I ran several real-world tests against other leading models and found it to be noticeably faster, more accurate, and significantly cheaper. Here’s what I discovered and how you can try it yourself.
What is GPT-5 Codex?
GPT-5 Codex is OpenAI’s latest model aimed specifically at software development. It promises faster reasoning, better accuracy on coding problems, and very low latency. The model uses a mixture-of-experts architecture and a 200k context window, which helps it handle large codebases and complex instructions.
How I accessed GPT-5 Codex
Instead of waiting for an official API integration in every tool, I used Kilo Code – an open-source AI agent that runs inside VS Code. Kilo Code lets you connect to various LLMs, and I configured it to route requests to GPT-5 Codex via OpenRouter. This gave me a direct way to test the model on real coding tasks without any custom setup.
- Install the Kilo Code extension in VS Code.
- Configure the extension to use OpenRouter as the provider.
- Select GPT-5 Codex from the available model list.
- Start a new coding session and give it a task.
Head-to-head coding tests
I put GPT-5 Codex up against GLM-4.6, Claude Sonnet 4.5, and Kimi K2 using a series of practical challenges. These included building a full-stack Next.js application, debugging a JavaScript snippet, generating a React component from a description, and creating a small AI-powered automation script.
Across the board, GPT-5 Codex produced working solutions more quickly and with fewer errors. Its responses felt snappier, and the code it wrote needed less manual tweaking. The tool-augmented reasoning also meant it could interact with the file system and terminal more reliably than the other models I tested.
Cost and performance
One of the biggest surprises was the cost. In my tests, GPT-5 Codex was about three times cheaper per request than the next best model, while delivering roughly twice the effective throughput. Lower latency and higher accuracy meant I spent less time waiting and less money on tokens overall.
For developers who rely on AI coding assistants daily, that combination of speed, quality, and price makes a real difference – especially on larger projects where context windows and reasoning depth matter.
Should you switch?
If you’re already using an AI code editor like Cursor, Bolt, or Replit Agent, adding GPT-5 Codex via Kilo Code is straightforward. It’s not about replacing your entire workflow, but about having a stronger model available for the toughest tasks. I’ll keep testing it as the tooling matures, but the early results are promising.
Tools used in this video
- GPT-5 Codex
- Kilo Code
- VS Code
- OpenRouter
- GLM-4.6
- Claude Sonnet 4.5
- Kimi K2
Questions
- How does GPT-5 Codex compare to Claude Sonnet 4.5?
- In my coding tests, GPT-5 Codex was faster, more accurate, and cheaper. It handled Next.js builds and debugging with fewer mistakes and lower latency.
- Can I use GPT-5 Codex for free?
- Access isn’t free, but it is significantly cheaper than comparable models – about three times less expensive per request in my testing. You can use it through OpenRouter with a pay-as-you-go plan.
- What tools do I need to start coding with GPT-5?
- You need VS Code and the Kilo Code extension. Configure it to use OpenRouter, select GPT-5 Codex, and you can start giving it coding tasks immediately.
- Is GPT-5 Codex only for web development?
- No, I tested it on full-stack JavaScript, React components, and automation scripts, but it’s a general-purpose coding model. It should work for any language or framework that the base model supports.
Video description and links
This article was written from the video's description with the help of AI. Watch the video for the full walkthrough.



