termdeck
Codex CLI guide: models and effort
Two things about Codex catch people out. The effort levels are a property of the model rather than a fixed list, and the context window it reports is larger than the one you actually get, because Codex reserves part of it for its own harness.
OpenAI
Each model advertises its own set of levels in the catalog. There is no fixed list, so a picker built from a hardcoded set will drop a level the day Codex adds one, run the turn at the default, and report the level you chose anyway.
Codex reserves part of the context window for its own harness. A model advertising 272,000 tokens gives you about 258,400, and one advertising 128,000 gives you about 121,600. A usage bar built on the advertised figure reads low by that margin.
Both are set when a turn starts. Plan mode is the exception worth knowing: it needs a specific resolved model rather than an unresolved default, so it will refuse rather than guess.
Codex rejects an effort level it does not recognise rather than quietly falling back. That is the failure worth having, because the alternative is a turn that runs at a different level than the one displayed everywhere in the interface.
The catalog is the authority on what exists. Reading it live is what keeps a new model or a new effort level from being invisible.
Compare
| Claude Code | Codex CLI | Grok CLI | |
|---|---|---|---|
| Pick a model | Yes | Yes | Yes |
| Reasoning effort | Yes | Yes, per model | Yes, per model |
| Separate thinking budget | Yes | No | No |
| Where the level list comes from | The engine | The model catalog | The initialize handshake |
| Reports its context window | Yes | Yes, larger than usable | Yes, and it is accurate |
| Auto compaction | Yes | Yes | At 85 percent of the window |
Termdeck drives all three from one console, so the differences above stop being three separate habits to remember.
FAQ
Codex advertises the full context window of the model but reserves part of it for its own harness. A model reporting 272,000 tokens leaves roughly 258,400 usable, so a bar built on the advertised number under-reports how full the context really is.
Whichever ones the model you picked advertises. The levels are a property of the model rather than a fixed list, so they change with the model and they change when Codex ships a new one.
It rejects a level the model does not advertise rather than silently substituting a default. That is deliberate: a silent fallback would run the turn at one level while every readout showed another.
Plan mode is requested with the model as part of the request, so an unresolved default cannot be sent. Pick a concrete model and plan mode will start.
Claude Code, Codex and Grok in one browser console