Agent workflows built on LLMs carry an unavoidable hidden dependency: the generational turnover of the underlying model. Codex's What's new lays out the timeline clearly — on September 21–25, GPT-6 Sol and GPT-6 Luna began rolling out in Codex at lower token prices than GPT-5.6; right after, GPT-5.5 is explicitly marked to retire from ChatGPT, ChatGPT Work, and Codex on October 14, 2026. This 'new model + scheduled retirement' pairing is not news to developers but a migration signal with a deadline: every prompt, eval baseline, and cost estimate still pinned to the old model must be validated and switched before the retirement window, or the workflow will silently get pricier or drift in behavior.
For Zero to Codex learners, this timeline is real-world material for 'model routing and selection': the course stresses that an agent's value shouldn't be pinned to one concrete model version — models retire, are replaced, and get repriced, so 'model choice' must be a replaceable configuration, not hardcoded. For Zero to Pro Agent Harness learners, 'model lifecycle management' must enter the harness design's must-answer list: how do you know which workflows depend on which model version? How do you run full regression before a model retires? Leave these unanswered and a seemingly stable agent system will silently fail on some ordinary day.