codex
**OpenAI Codex** is the **pioneering code generation model that powered the original GitHub Copilot, fine-tuned from GPT-3 on billions of lines of public code from GitHub** — proving for the first time that large language models specialized for code could provide practical, real-time coding assistance in IDEs, creating the "AI coding" category that now includes Copilot, Cursor, Tabnine, and dozens of competitors, before being deprecated in March 2023 as its capabilities were absorbed into GPT-3.5 and GPT-4.
**What Was Codex?**
- **Definition**: A family of GPT-3-descendant models fine-tuned on publicly available code from GitHub — available as `code-davinci-002` (12B parameters, most capable) and `code-cushman-001` (smaller, faster), exposed through OpenAI's API for code generation, completion, and translation tasks.
- **The Original Copilot**: GitHub Copilot (launched June 2021) was powered entirely by Codex — the model that first demonstrated that AI autocomplete in IDEs was not just possible but genuinely useful for everyday programming.
- **Deprecation (March 2023)**: OpenAI deprecated the Codex API as GPT-3.5 and GPT-4 absorbed and exceeded its code generation capabilities — code generation became a standard feature of general-purpose models rather than requiring a specialized model.
**Codex Capabilities**
| Capability | How It Worked | Impact |
|------------|------------|--------|
| **Code Completion** | Predict next lines from context | First practical AI autocomplete |
| **Natural Language to Code** | "Sort this list by date" → code | Democratized coding for non-experts |
| **Code Translation** | Python → JavaScript conversion | Cross-language development |
| **Code Explanation** | Code → natural language description | Code comprehension aid |
| **Bug Detection** | Identify issues from context | Early AI-assisted debugging |
**Performance Benchmarks**
| Benchmark | Codex (code-davinci-002) | GPT-3 (text-davinci-002) | GPT-4 (successor) |
|-----------|------------------------|------------------------|-------------------|
| HumanEval (Python) | 47.0% | 0% | 67.0% |
| MBPP (Python) | 58.1% | ~10% | 83.0% |
| Languages supported | 12+ | Code not primary | All major languages |
**Legacy and Impact**
- **Created the AI Coding Category**: Before Codex/Copilot, AI code assistance was an academic curiosity. Codex made it a practical, daily-use tool for millions of developers.
- **Proved Specialization Works**: Demonstrated that fine-tuning a general LLM on domain data (code) dramatically improves domain performance — a lesson applied to medical (Med-PaLM), legal (Legal-BERT), and financial (BloombergGPT) AI.
- **$100M+ Business**: Copilot (powered by Codex) became GitHub's fastest-growing product, reaching millions of paid subscribers and proving the commercial viability of AI developer tools.
- **Deprecated but Absorbed**: Codex's capabilities weren't lost — they were integrated into GPT-3.5 and GPT-4, which now handle code generation as a standard capability alongside natural language understanding.
**OpenAI Codex is the model that launched the AI coding revolution** — proving that LLMs fine-tuned on code could provide practical, real-time development assistance and creating a multi-billion dollar market for AI coding tools that fundamentally changed how software is written.