2.4 Choosing Models & Reasoning Effort

Not every task deserves the same amount of thinking, and Codex gives you real control over that dial instead of a single one-size-fits-all setting.

On models: OpenAI ships a family of Codex-optimized models rather than a single fixed one, and the lineup keeps advancing — GPT-5-Codex was the first version purpose-built for agentic software engineering rather than adapted from a general chat model, GPT-5.1-Codex-Max extended that with stronger long-horizon reasoning and context compaction for sessions that run for hours, and GPT-5.2-Codex pushed further on large-scale refactors, native Windows support, and reading screenshots and technical diagrams directly (more on that in Module 4). The specific name matters less than the pattern: check /status or your Codex settings periodically to see which model your session is actually running, because the newest one isn't always the default for every account on day one, and it's worth deliberately opting in rather than assuming you're already on it.

On reasoning effort: within whichever model you're running, you can dial how hard it thinks about a given task. A quick, obvious fix doesn't need maximum reasoning effort — it's slower and costs more for a decision that was never actually hard. A subtle bug, a genuine architectural tradeoff, or a large refactor spanning many files is exactly where turning the effort up earns its cost. Running everything at maximum effort out of habit is a common beginner pattern, and it's pure waste on the easy ninety percent of tasks that didn't need it.

The practical rule: match the setting to how expensive being wrong would be, not to how the task feels in the moment. A five-line config change and a payment-flow refactor should not get the same treatment, even if both technically "just need a quick look."