Only gpt-5.3-codex Publishes Reasoning Effort

The mistake

Passing a reasoning-effort parameter to a model other than the Codex-optimized one, assuming every current model publishes the same set of effort levels simply because one of them does and the concept feels like it should generalize across a model family.

Why this happens

Reasoning-effort control is a genuinely useful, well-documented feature on the one model that publishes it, and it's natural to assume a capability that prominent gets treated the same way across the rest of the lineup. It doesn't: no other current model's own reference page states any reasoning-effort levels at all, published or otherwise. This site records that honestly as "not published" for every other model, specifically rather than assuming the Codex-optimized model's four levels apply universally just because they're well documented in one place.

Why it matters

Code or configuration that assumes every model accepts the same effort parameter risks either an outright error on models that don't support it, or a silently ignored setting that changes nothing about how the request is actually processed — both worse outcomes than knowing up front which single model this control genuinely applies to. Planning a workload's behavior around reasoning-effort tuning only works on the one model where that control is real.

The fix

Treat reasoning-effort control as specific to the Codex-optimized model until another model's own page states otherwise — don't infer support from a general sense that "newer models probably support this too." If your workload specifically wants that control, that's a real, additional reason to choose the Codex-optimized model over an alternative, on top of whatever its price and context-window tradeoffs already suggest.

It's also worth revisiting this periodically rather than assuming it's permanent — a capability documented for one model today is a reasonable candidate to appear on others in a future release, and this site's weekly fact-check is specifically what would catch that the moment it happens, rather than leaving the old "Codex-only" assumption to go stale unnoticed on a page nobody rechecks.

See also

Reasoning Effort, and Why It's Codex-Only For Now lays out the four published levels and what each is documented to do.