Unsuffixed GPT-5.6 Is Just an Alias for Sol

The mistake

Specifying the unsuffixed gpt-5.6 id in a config or API call and expecting it to resolve to some auto-selected or "latest" variant — then being confused when its behavior, price and specs exactly match one named model.

Why this happens

The unsuffixed name reads like "whichever GPT-5.6 variant is appropriate," the way some vendors' unversioned aliases route dynamically. That's not what happens here. OpenAI documents gpt-5.6 as a fixed alias that routes to GPT-5.6 Sol — Sol "roughly corresponds to the unsuffixed model tier used in earlier GPT-5 families." It is a pointer to one model, not a router across the family.

Why it matters

Three ways. A cost comparison that lists gpt-5.6 beside GPT-5.6 Sol is comparing a model with itself. A config that says gpt-5.6 is running Sol at Sol's prices — which are currently promotional, so that budget has an expiry date it doesn't show. And the alias is not a "latest model" pointer: it did not move when the GPT-6 models arrived. The Codex CLI's banner now names GPT-6.1 Sol, a different model entirely, so a config pinned to gpt-5.6 quietly diverged from the CLI's default the day the default changed.

The fix

Use explicit model ids in configuration and scripts — gpt-5.6-sol rather than gpt-5.6, so the file says plainly what runs. Treat the alias and its target as identical for every pricing and capability purpose. And when you want the current Codex default, set it deliberately rather than relying on an alias to follow it; no published gpt-6 alias exists to do that for you.

Finding the alias in your setup

Search configs, scripts and environment variables for gpt-5.6 where it isn't followed by -sol, -terra or -luna; the bare form is easy to read past because it looks complete. In the Codex CLI, /status shows the model a session is actually running — if it reports GPT-5.6 Sol where you expected the current default, an alias or a pinned id is the reason.

When you replace it, decide deliberately. Pinning gpt-5.6-sol changes nothing about cost or behavior; it only makes the file honest. Moving to a GPT-6 model changes both, and deserves a test on real work first. As a general rule, prefer full ids to aliases in anything checked into a repository: a model change then arrives as a reviewed diff instead of a silent difference between what the file says and what runs.