Technical Context
I went to check what exactly happened in Codex, and the picture is double-edged. On one hand, on July 12, OpenAI temporarily removed the 5-hour limit for Plus, Pro, and Business. On the other hand, a very nasty bug surfaced in Terra: the agent, in the middle of work, can stop writing code and start just answering like a chat.
For those building AI automation around code generation, this is not cosmetic. I would consider this a change in the operational behavior of the system, not just UI news. The limit was lifted temporarily because after GPT-5.6 Sol, load spiked sharply, but the weekly cap hasn't gone anywhere, and the pool is still shared with other ChatGPT Work tasks.
By symptoms, the bug looks like this: the session starts normally, model Terra, effort high, speed standard, and then the agent suddenly stops acting as an agent. In discussions, people initially suspected /plan, but in real cases, a new session with the same parameters restored normal behavior. I look at such things as an engineer: if a new session fixes the problem, then the failure is most likely in the state of that specific session, not in the prompt itself.
And that's where I hit pause. When a code development tool starts switching modes mid-task, it's no longer a convenience issue but a question of pipeline predictability.
Impact on Business and Automation
Who wins? Teams that were hitting the old 5-hour ceiling with long agent runs. Now you can run refactoring, test generation, and technical migrations longer without constant pauses.
Who loses? Those who already tied a production process to a single long Terra session without a safety net. If the agent suddenly goes into chat mode, the pipeline hangs, and the developer starts manually digging through context.
I would right now implement three things: short atomic tasks instead of one endless one, automatic restart of a new session on mode change, and distributing critical scenarios across models. This is no longer AI integration theory, but standard engineering hygiene.
At Nahornyi AI Lab, we fix exactly these spots for clients: we don’t just plug in a model; we build a resilient AI architecture with retries, fallback logic, and proper cost control. If Codex is already affecting your release timelines, we can review your process together and build AI automation so that one session's whim doesn't break the entire development.