Codex CLI Defaults to Sol, Not the Codex Model

The mistake

Assuming that because a model is literally named for Codex and described by OpenAI as "optimized for agentic coding tasks in Codex or similar environments," the Codex CLI must default to it. It doesn't. Run the CLI with no model flag at all and check the banner it prints at startup — the model actually answering is the general frontier model, not the Codex-named one.

Why this happens

The naming genuinely suggests the opposite of what's true, and nothing in the CLI's first-run experience corrects the assumption unless you go looking for it. The unsuffixed model name is documented as an alias that routes to the frontier model too, which reinforces the same wrong intuition from a different angle — everything about the naming points toward the Codex-branded model being the "real" default, and it simply isn't.

Why it matters

The two models genuinely differ on cost, context window, and how recent their training data is — not on quality in some vague sense, on three concrete, budgetable axes. If you're planning spend or context-window headroom against the wrong one of the two, the numbers you plan against will be wrong in a specific, predictable direction: the actual default costs more per token and carries a noticeably larger context window than the Codex-branded alternative, and its knowledge is more recent. None of that is a problem on its own — it's simply not what most people assume they're getting when they see "Codex" in a model's name.

The fix

Check the model your CLI is actually running with before budgeting anything, either from the startup banner or with the /status slash command inside a session. If the default is genuinely what you want — and for most agentic coding work, its larger context window is a real advantage worth the higher per-token price — nothing needs to change. If you specifically want the Codex-optimized model's lower price and its published reasoning-effort control, switch to it explicitly with /model rather than assuming the name gets you there by default.

See also

The full side-by-side is worth reading in full: it's the single most useful comparison this site can offer, because no OpenAI page states the trade-off between the two models directly.