One Name, Two Products, Zero Overlap: The Codex Identity Crisis

The Name Codex Now Means Two Completely Different Things—and That's a Problem

There's a naming disaster buried in OpenAI's engineering portfolio, and it's worth understanding if you're evaluating code automation tools: the word "Codex" refers to two entirely separate products, separated by a complete shutdown and a two-year gap, built on completely different architectures, and serving completely different use cases.

The confusion is real. The original Codex launched in August 2021 as a code-completion API through the completions endpoint. It was eventually shut down, and OpenAI reused the name in 2025 for a cloud-based agentic coding system that runs in sandboxed environments. They're not related. They don't use the same models. They don't share the same pricing model. And they definitely don't solve the same problem.

Product One: The Original Codex (2021–2023)

The original Codex lived an unglamorous life as a limited beta. It launched on August 10, 2021, in private beta and never left that status. OpenAI's 2023 deprecation email cited advancements in GPT-3.5 models for coding tasks as the reason for sunsetting it.

What did it do? Simple: It turned English into code through the completions API. Hand it a natural-language description, and it would write a code snippet. That was the entire scope. It lived as a code-completion tool in a landscape already dominated by GitHub Copilot, which shipped in October 2021 using the same underlying Codex model.

The problem: some teams ran production traffic on what was explicitly labeled a limited beta, so when OpenAI shut it down around March 2023, there was fallout. But from a technical standpoint, OpenAI's decision was straightforward—newer models did the same job better.

Product Two: The New Codex (2025–Present)

The 2025 version is an entirely different animal. It launched as a cloud-based software engineering agent available to ChatGPT Pro, Business, and Enterprise users, running research preview starting May 2025. This Codex doesn't just complete code—it performs complex end-to-end tasks such as writing features, answering questions about your codebase, fixing bugs, and proposing pull requests for review, with each task running in its own cloud sandbox environment preloaded with your repository.

The 2025 Codex agent is anchored on codex-1, an OpenAI o3-family model specialized for coding tasks through additional reinforcement-learning training on real-world software-engineering trajectories. This is not a completion tool. It runs in cloud-hosted, isolated containers preloaded with the user's repository code, can read and edit files, run terminal commands, and execute tests inside the sandbox.

The workflow is fundamentally different. The May 2025 launch positioned Codex as a "delegate-and-review" workflow rather than a real-time pair-programmer, which is the niche occupied by GitHub Copilot and Cursor. You write a high-level task description, hand it off to the agent, monitor progress in real time, and review the output before merging. Task completion typically takes between 1 and 30 minutes, depending on complexity.

Why the Name Reuse? Organizational Consequences

Reusing the name "Codex" creates genuine friction for buyers and engineers trying to understand what OpenAI is actually shipping. If a team member says "we should look at Codex," half the conversation is now spent clarifying which Codex—the deprecated 2021 API or the 2025 cloud platform. The deprecation documentation is sparse; OpenAI's 2021 launch post still carries a line stating "This launch post is outdated," with no pointer to what happened next.

The pricing tells a different story too. Codex is included at no additional cost in ChatGPT Plus ($20/month), Pro ($200/month), Team (from $25/seat/month annual), Enterprise, and Edu plans. But usage limits and longer-term pricing structure may evolve as the product moves out of initial release; through 2026 the practical bottleneck has been compute-time fair-use limits inside ChatGPT Plus rather than feature gates. The new Codex is compute-intensive and cloud-based; the old one was an API call.

A standalone Codex CLI (open-source, MIT) provides terminal access to the same model and tooling pattern for users who prefer command-line workflows. A smaller version of codex-1, a version of o4-mini designed specifically for use in Codex CLI, supports faster workflows in the CLI and is optimized for low-latency code Q&A and editing.

What This Means for Your Team

If you have legacy code referencing the original Codex API from 2021–2023, you've already migrated (or you're running on GPT-3.5 or GPT-4). That decision is made.

If you're evaluating the new Codex as an agentic tool for software engineering, understand what it actually is: a cloud-based task executor that operates on sandboxed copies of your repository, runs for 1–30 minutes per task, and operates in a delegate-and-review workflow. It's not a pair-programmer. It's not inline code completion. It's not a chat interface with code features bolted on. It's a fundamentally different category from GitHub Copilot, Cursor, or traditional code-completion tools.

The organizational prerequisite is straightforward: you need code already in a git repository, a willingness to let an agent stage changes for review, and tolerance for compute time that varies by task complexity. For organizations already comfortable with CI/CD automation and code review processes, the mental model should be familiar. For smaller teams or individual developers used to real-time pair-programming, the workflow is a step change.

The naming confusion isn't fatal, but it's worth clarifying with your team before evaluation. Call the 2021–2023 thing "legacy Codex" or just not at all—it's gone. The 2025 cloud agent is what OpenAI is shipping now. They're different products with a shared name, and pretending they're the same thing will only slow down your decision-making.

Dimension	Original Codex (2021–2023)	New Codex (2025–Present)
Primary Use	Code completion from natural language	End-to-end task execution in repository
Architecture	Completions API endpoint	Cloud-based agentic platform with sandboxes
Workflow	Real-time inline suggestions	Delegate, monitor, review (1–30 min per task)
Status	Deprecated March 2023	Active research preview, rolling availability
Access	API (now removed)	ChatGPT Pro/Business/Enterprise + CLI
Pricing	Pay-per-API-call	Included in subscription; compute limits apply
Competitive Set	GitHub Copilot, IntelliCode	AI coding agents, CI/CD automation, software engineering orchestration