GPT-5.5 is in Codex before the API, and that changes how builders should read this launch

GPT-5.5 is out, and the early verdict is pretty simple: it looks fast, capable, and good at the kinds of hands-on tasks people actually throw at frontier models now.

But the detail worth paying attention to is that GPT-5.5 landed in OpenAI Codex and for paid ChatGPT users before it landed in the public API. That is not just a rollout footnote. It is a clue about where OpenAI thinks the highest-value usage is happening.

What happened

Simon Willison’s write-up points to a semi-official path for using GPT-5.5 through Codex-linked tooling even before the API arrives. OpenAI appears comfortable with developers accessing the model through the same subscription-connected mechanisms used by Codex, and that opens the door for tools like the new llm-openai-via-codex plugin.

In plain English: if you are already working inside agent and terminal workflows, GPT-5.5 may be testable there first. That makes this release feel less like a normal API launch and more like a signal that the agent stack is now part of the default product surface.

• Codex and paid ChatGPT got GPT-5.5 first
• public API access is still pending
• subscription-linked tooling is becoming a practical access path
• reasoning effort settings can materially change output quality and token use

Why it matters in practice

If you build with models for a living, this changes the first question from "what does the benchmark chart say?" to "where can my team actually use this tomorrow?" Right now the answer may be agents, terminals, and subscription-backed workflows before standard API integrations catch up.

That also means evaluation gets messier. Willison highlights that higher reasoning settings improved results dramatically on his pelican test, but they also burned far more reasoning tokens. So the next real builder question is not whether GPT-5.5 is better. It is whether it is better enough at your task mix to justify the cost once API pricing lands.

Who should care: teams already using Codex-style workflows, indie developers trying to squeeze more value out of subscriptions, and anyone deciding whether to standardize on GPT-5.4, wait for API access, or start prototyping against this new path now. The practical move is to test the workflow, not just the model.