You access GPT 5.3 Codex today primarily through the Codex experiences tied to ChatGPT subscriptions and through GitHub Copilot where it’s rolling out as a model option. OpenAI states that anyone with a supported ChatGPT subscription can use Codex across the CLI, web, IDE extension, and the Codex macOS app using their ChatGPT login, with usage included in the subscription (and optional extra credits). That’s described directly in the Codex app announcement: Codex app availability & pricing. Separately, GitHub’s changelog states GPT-5.3-Codex is “now rolling out” in GitHub Copilot, which is the other major access route for many developers: GitHub Copilot GA.
A practical “access checklist” you can follow looks like this:
If you’re a ChatGPT subscriber (Plus/Pro/Business/Enterprise/Edu):
Install and use the Codex macOS app if you want agent threads and diff-centric workflows: Codex app.
Use Codex through supported CLI / web / IDE-extension surfaces with your ChatGPT login (per OpenAI’s availability statement in the same post).
If you’re a GitHub Copilot user:
- Look for GPT 5.3 Codex in Copilot’s model selection (rollout and admin policy may apply for org accounts): GitHub Copilot GA.
If you need API-key workflows:
- Pay attention to OpenAI’s Codex developer changelog; it notes that some related variants (like Codex-Spark) may not be available via API at launch, and API availability can be phased: Codex changelog.
This matters because “available in the app” and “available in the API” are not always the same thing on day one.
If your end goal is to use GPT 5.3 Codex inside a documentation site or developer portal (like Milvus.io), the most realistic way to “access it” is to pick one surface first (Codex app or Copilot) for evaluation, then design your production workflow around retrieval and verification. For example, index your docs and troubleshooting guides into Milvus or Zilliz Cloud, retrieve the most relevant passages for a user question, and pass them into GPT 5.3 Codex with a strict instruction to answer using only retrieved context. That lets you evaluate the model in the same constraints you’ll ship: grounded answers, predictable output formats, and measurable quality.