OpenAI Killed Codex in 2023. Then They Brought It Back. Here's What That Tells Us.

OpenAI Killed Codex in 2023. Then They Brought It Back. Here's What That Tells Us.

March 2023. OpenAI deprecated the Codex API. The model that had powered GitHub Copilot, that had proved to millions of developers that AI could actually write useful code, was unceremoniously retired. Migrate to GPT-3.5-Turbo and GPT-4, they said. Codex was done.

Then in May 2025, OpenAI launched a new product called... Codex. A cloud-based coding agent built into ChatGPT.

Same name. Completely different animal. And the gap between those two products tells you everything about where AI-assisted development is actually going.

The Original Codex: A Model That Proved a Market

Let's rewind. OpenAI introduced Codex in August 2021 as a specialized model fine-tuned on publicly available code. It descended from GPT-3, trained specifically to translate natural language into programming languages. The pitch was simple: describe what you want, get code back.

The real genius wasn't the model itself. It was what Microsoft and GitHub did with it. GitHub Copilot launched as a technical preview in June 2021, powered by Codex, and it did something no standalone AI coding tool had managed before. It met developers where they already were: inside their editor.

By mid-2024, GitHub Copilot had crossed 1.8 million paid subscribers. That number matters because it validated an entire category. Before Copilot, "AI writes code" was a demo you'd see at a conference and forget about. After Copilot, it was a line item in every engineering team's budget.

But here's the thing nobody talks about: Codex the model was already being outpaced before it was deprecated. In February 2023, just one month before the shutdown announcement, GitHub's own blog detailed upgrading Copilot to a newer, improved Codex variant. The original was already obsolete. The child had outgrown the parent.

When OpenAI officially pulled the plug on March 23, 2023, less than two years after launch, the stated reason was blunt: "Our latest models are now our best models for coding-related tasks." GPT-3.5-Turbo and GPT-4 could do everything Codex did, plus handle the broader context that real coding requires. Understanding documentation. Parsing error messages. Reasoning about architecture. Codex was a specialist in a world that had started demanding generalists.

The Death of the Specialist Model

Codex's deprecation wasn't just a product decision. It was a signal about how the entire AI industry would evolve.

In 2021, the assumption was that we'd need purpose-built models for different domains. A model for code. A model for text. A model for images. Each fine-tuned, each specialized, each requiring its own API and integration work. Codex was the poster child for this approach.

What actually happened was that general-purpose models got good enough at everything to make specialists unnecessary. GPT-4 didn't just match Codex at code generation. It surpassed it, because it could bring world knowledge and reasoning capabilities that a code-only model couldn't touch. When a developer asks "write me a function that calculates compound interest," the model that understands finance and code produces better output than the model that only understands code. Every time.

I've seen this pattern play out repeatedly in software. The specialized tool that does one thing brilliantly gets absorbed by the platform that does everything well enough. Remember when we had separate apps for todo lists, note-taking, project management, and wikis? Now it's all Notion. Same consolidation happened with AI models, just at warp speed.

The market followed suit. Amazon launched CodeWhisperer (now Amazon Q Developer). Google shipped code generation in Gemini. Every major AI lab realized the same thing: you don't need a code model. You need a good model that can also code.

The specialist era of AI lasted roughly 18 months. That's how fast the ground shifted under an entire product category.

Why Integration Beats Standalone Every Time

There's a deeper lesson in the Codex story that goes beyond model architecture. It's about distribution.

Codex as a standalone API had an adoption problem. Developers had to build their own integrations, manage their own prompt engineering, handle their own context windows. Powerful but raw. A building block, not a product.

GitHub Copilot was a product. It lived in VS Code. It understood your open files. It autocompleted as you typed. The friction was near zero.

This is why the IDE-integrated approach won so decisively. Copilot, Cursor, Cody by Sourcegraph, Tabnine. The winners in AI coding aren't the ones with the best model. They're the ones with the tightest integration into the developer's existing workflow. The model is a commodity. The context is the product.

I've shipped enough features to know that the best technology almost never wins on its own. Distribution wins. Convenience wins. The tool that's already open on your screen when you need it wins. Codex as an API required developers to change their workflow. Copilot slid into the workflow they already had. That's the whole ballgame.

The Resurrection: Codex 2025 Is a Different Beast

In May 2025, OpenAI brought the Codex name back. But the new Codex isn't a fine-tuned code model you hit with API calls. It's a cloud-based coding agent living directly inside ChatGPT. It reads repositories, writes code across multiple files, runs tests, and opens pull requests. It operates in a sandboxed cloud environment, acting as an autonomous junior developer that works asynchronously.

This isn't a resurrection. It's a reincarnation.

The original Codex said: "Give me a prompt, I'll give you code." The 2025 Codex says: "Give me a task, I'll do the engineering work." We've gone from code generation to code execution. From autocomplete to agent. That's not an incremental improvement. It's a different product category entirely.

And here's what I find genuinely fascinating about this move: it validates the exact lesson that killing the original Codex taught OpenAI. The new Codex is deeply integrated into ChatGPT. Not a standalone API. Not a separate product you have to go find and figure out. It's embedded in the tool millions of people already use every day.

OpenAI learned from their own history. The original Codex failed as a standalone API because developers needed products, not primitives. The new Codex succeeds (or at least has a shot at succeeding) because it's a product built on top of the primitives that replaced the original.

Now, the competitive angle here is interesting. GitHub Copilot, which the original Codex birthed, is in a weird position. OpenAI, its model provider, is offering a competing product under the same brand name that started it all. Microsoft owns both GitHub and a massive stake in OpenAI, so this is more family argument than competitor war. But the strategic tension is real, and I don't think it's resolved yet.

So What Does This Mean for You?

The two-act story of Codex has three takeaways I think every developer and engineering leader should internalize.

The model layer is commoditizing fast. Codex was special in 2021. By 2023, three different models could replace it. By 2025, the model isn't even the point. The agent layer on top is where the value lives. If your AI coding strategy is "pick the best model," you're optimizing the wrong variable.

Where an AI tool lives in your workflow matters more than how smart it is. The original Codex API lost to IDE-integrated tools. The new Codex is betting on ChatGPT as its integration surface instead. When you're evaluating AI coding tools, ask "how does this fit into my existing process?" before you ask "how good is the model?"

The direction is clear: autocomplete to autonomous agents. Codex 2021 generated code snippets. Copilot 2022 autocompleted lines. Codex 2025 executes multi-file tasks asynchronously. Each iteration gives the AI more autonomy and more scope. The next step isn't better autocomplete. It's AI that handles entire workstreams while you focus on architecture and product decisions.

I think we're about 18 months away from AI coding agents that can reliably handle most routine implementation work. Bug fixes, test writing, straightforward feature additions, migrations. Not the creative architectural work. Not the ambiguous product decisions. But the mechanical translation of clear requirements into working code.

If your entire value as an engineer is writing CRUD endpoints, the Codex story should keep you up at night. If your value is in system design, understanding tradeoffs, and making judgment calls under ambiguity, this is the most exciting time to be in the industry.

Codex died and came back as something far more ambitious. Your career strategy should do the same.

Photo by Rahul Viswanath on Unsplash.