Agentic Coding Arena — Frontier AI Coding Comparison

This comparison was conducted by Logic, Inc. — The Language-Optimized General Intelligence Corp.

This repository contains a multi-model comparison of demo applications generated by leading frontier AI coding models, using identical prompts for each example.

Participating models include OpenAI's GPT-5 family, Anthropic's Claude family (Opus, Sonnet, Haiku), Google's Gemini family, and Moonshot AI's Kimi family, with more models added over time.

This is a fork of OpenAI's GPT-5 Coding Examples repository, extended to include multiple model implementations of the same prompts for comparison purposes.

Each model was given the same natural language descriptions to scaffold websites, front-end applications, games, and interactive UIs — allowing for a direct comparison of their respective coding capabilities. All applications were generated in a single one-shot prompt, without any manual code editing.

Repository Structure

• public/apps/{model-id}/ - Applications generated by each model (e.g., public/apps/gpt-5/, public/apps/opus-4.7/, public/apps/kimi-k2.6/)
• /examples - YAML files containing the prompts used for all models
• /app - Next.js App Router frontend
• /components - React components for the comparison UI
• /scripts - Automated generation pipeline

Explore Examples

You can explore the demos by cloning this repo and running it locally.

cd front-end
npm install
npm run dev

The app will be available at localhost:3000.

From there, you can view examples from all models side-by-side, see the prompts that created the code, and compare the different approaches taken by each AI model.

Generate Apps

The generation pipeline keeps the model owner, API host, and coding harness as separate configuration fields. For example, Qwen3.8 27B is generated by the OpenCode harness through OpenRouter.

Install OpenCode and authenticate OpenRouter with opencode auth login, or export OPENROUTER_API_KEY. Then run:

npm run generate -- qwen-3.8-27b

OpenCode receives the fully qualified model ID openrouter/qwen/qwen3.8-27b. The model is included in the public comparison and stats once its generated apps are present.

Build Your Own

You can experiment with these same prompts using any of the models:

Using GPT-5 / GPT-5.1

• ChatGPT – Choose GPT-5 or GPT-5.1 to generate and preview code in the browser
• Your favorite IDE or coding tool – Use GPT-5/5.1 within your existing workflow to generate and refine code
• API – Access GPT-5/5.1 through OpenAI's API for programmatic generation

Using GPT-5.5

• ChatGPT – Choose GPT-5.5 to generate and preview code in the browser
• Your favorite IDE or coding tool – Use GPT-5.5 within your existing workflow to generate and refine code
• API – Access GPT-5.5 through OpenAI's API for programmatic generation

Using Claude Opus 4.1

• Claude – Use Claude Opus 4.1 in the browser interface
• Claude Code – A lightweight coding agent that runs in your terminal
• API – Access Claude Opus 4.1 through Anthropic's API for programmatic generation

Using Claude Opus 4.5

• Claude – Use Claude Opus 4.5 in the browser interface
• Claude Code – A lightweight coding agent that runs in your terminal
• API – Access Claude Opus 4.5 through Anthropic's API for programmatic generation

Using Claude Opus 4.7

• Claude – Use Claude Opus 4.7 in the browser interface
• Claude Code – A lightweight coding agent that runs in your terminal
• API – Access Claude Opus 4.7 through Anthropic's API for programmatic generation

Using Claude Opus 4.8

• Claude – Use Claude Opus 4.8 in the browser interface
• Claude Code – A lightweight coding agent that runs in your terminal
• API – Access Claude Opus 4.8 through Anthropic's API for programmatic generation

Using Claude Sonnet 4.5

• Claude – Use Claude Sonnet 4.5 in the browser interface
• Claude Code – A lightweight coding agent that runs in your terminal
• API – Access Claude Sonnet 4.5 through Anthropic's API for programmatic generation

Using Gemini 3

• Google AI Studio – Use Gemini 3 in the browser interface
• Gemini – Chat with Gemini 3
• API – Access Gemini 3 through Google's API for programmatic generation

Using Gemini 3.5 Flash

• Google AI Studio – Use Gemini 3.5 Flash in the browser interface
• Gemini – Chat with Gemini 3.5 Flash
• API – Access Gemini 3.5 Flash through Google's API for programmatic generation

Using Kimi K2.6

• Moonshot AI – Use Kimi K2.6 in the browser interface
• API – Access Kimi K2.6 through Moonshot AI's API or via Fireworks AI

Choose an example from the /examples directory, copy its prompt for inspiration, and adapt it to your own needs. All models can scaffold complete applications from natural language descriptions.

Comparison Highlights

All models demonstrate strong capabilities in:

• Single-file HTML/CSS/JavaScript applications
• Interactive games and animations
• Data visualization and dashboards
• Responsive design and modern UI patterns
• Complex state management and user interactions

The multi-model comparison reveals interesting differences in:

• Code organization and structure
• Design aesthetics and UI choices
• Implementation approaches for the same functionality
• Performance optimizations
• Error handling strategies

Note

This repository is for comparison and research purposes. All applications were generated in a single prompt without manual editing, showcasing the raw capabilities of each model.