What happened
This article covers the product and allowance changes that followed the GPT-5.6 model-family launch. It is not a second introduction to the models.
Less than a day after the launch, OpenAI said it was changing parts of the ChatGPT Work and Codex experience. In a follow-up post, Tibo Sottiaux, who works on Codex and ChatGPT, said the team had reset usage limits, would reset them again later that day, and was shipping an initial set of fixes.
The changes address several kinds of launch friction. OpenAI made high-compute settings too easy to use without clearly showing their effect on limits, reorganized the desktop app in ways that made chats and projects harder to find, and introduced regressions in some multi-agent workflows and plugins. The update also explicitly said Codex is not going away as ChatGPT Work develops.
OpenAI said it was changing defaults and the model picker so users were less likely to choose unnecessarily expensive settings. It also promised clearer usage and reset timing, a more familiar and customizable sidebar, better communication about when to use ChatGPT Work versus Codex, and further fixes the following week.
Why it matters
This is a useful reminder that a model launch changes more than model quality. It changes the surrounding interface, defaults, usage budget, and assumptions about how people work. A more capable setting is not automatically an improvement if users can select it without understanding how quickly it consumes their allowance.
The Codex clarification matters too. ChatGPT Work and Codex may share more of the same workspace over time, but builders still need to know which product owns their files, sessions, limits, and agent workflows. Product convergence can reduce friction for some users while making the boundaries harder to understand for others.
The timing also makes usage limits a competitive product issue. Anthropic reset Claude’s limits the day before, extending access just as users were deciding whether to stay with an existing coding workflow or try GPT-5.6. In both cases, capacity and transparency became part of the launch story.
Who should care
Codex and ChatGPT Work users should pay attention if they use high-compute reasoning, multi-agent tasks, plugins, or the desktop app. Builders comparing agent products should also evaluate the entire workflow around a model: defaults, usage visibility, reset behavior, and how gracefully the product handles regressions.
What builders should do next
Check which reasoning or compute setting your tool is using before starting a long task, and look at the usage panel rather than assuming the default is the cheapest sensible option. Keep a fallback workflow available while the promised fixes land.
When testing GPT-5.6 against another agent, compare more than the final answer. Record how much allowance the task consumes, whether the agent recovers from a failed subtask, and how easy it is to understand what will happen before you approve it.
The practical takeaway is that OpenAI’s response is a meaningful product signal: ambitious agent launches need clear cost controls and stable workflows on day one, not only a stronger model underneath.
End of article