OpenAI Cracks Open Codex, the Brain Behind GitHub Copilot

OpenAI just put a name and an API key on the thing that’s been quietly powering GitHub Copilot: Codex. As of today it’s live in private beta, and if you’ve been following the Copilot debate over the last few weeks, this is the missing piece — the model doing the actual “understanding” when Copilot turns your comment into working code.

Codex is a descendant of GPT-3, but instead of being a generalist trained mostly on internet text, it’s been fine-tuned specifically on code — 159GB of public Python pulled from 54 million GitHub repositories. That’s a genuinely huge slice of the open-source world poured into one model, and it shows in the numbers OpenAI is sharing: Codex reportedly solves more than 70% of the problems on OpenAI’s own HumanEval benchmark, a set of hand-written programming problems designed to test whether a model can produce correct, runnable code from a natural-language description. Base GPT-3, for comparison, solves essentially none of them. That’s not an incremental improvement — that’s a different category of tool.

What’s notable is that this isn’t a research paper or a demo video. It’s an API. Private beta access is limited for now, but the intent is clear: OpenAI wants developers building products on top of this, not just marveling at it in a blog post. Copilot has been the flashy public face since its technical preview launched in VS Code back in June, but Codex-the-API opens the door to a much wider set of applications — natural-language-to-SQL tools, code explanation assistants, automated test generation, chatbots that can write and execute small scripts on the fly. Basically anything where “describe what you want” could replace “write it by hand.”

I’ll be honest, the 70% HumanEval number is the kind of stat that’s easy to over-read. Solving a benchmark problem isn’t the same as writing production-quality code that fits your team’s conventions, handles edge cases, and doesn’t quietly reproduce someone else’s copyrighted snippet — which is exactly the concern critics have been raising about Copilot since its preview dropped. Codex being trained on public repos doesn’t make that concern go away; if anything, opening it up as a general-purpose API means more products downstream will inherit the same licensing ambiguity.

Still, it’s hard to overstate how fast this space is moving. Six weeks ago, “AI pair programmer” was a novel pitch attached to one product. Today it’s an API anyone can theoretically build on. If Codex access widens beyond the current beta, expect a wave of code-adjacent tools to show up trying to ride this wave — some genuinely useful, some just Copilot with a different UI slapped on top. The interesting question isn’t whether AI can autocomplete a function anymore. It’s what happens once every developer tool has one of these bolted on.