Models and effort
How the model picker names models, how the effort slider fits each model, and how chats remember both.
Every chat has a model and, where the model supports it, a reasoning effort. Both are picked in the composer. The model picked before the first message also decides which harness runs the chat.
Model names
The picker shows models by name instead of by id. The id stays what the harness receives. When Codex gives a model a display name, the picker uses it; otherwise the app turns the id into a name:
| Id | Shown as |
|---|---|
claude-opus-5-5 | Claude Opus 5.5 |
claude-sonnet-4-20250514 | Claude Sonnet 4 |
gpt-5.5 | GPT-5.5 |
anthropic/claude-haiku-4-5 | Claude Haiku 4.5 |
The rules behind the names:
- A provider prefix before
/is dropped, and so is a trailing date such as20250514. - Claude ids read as "Claude", the family and the version. GPT ids read as "GPT-" and the version.
- Well-known families keep their own spelling: GLM, DeepSeek, xAI, Qwen, Llama, Mistral, Gemini, Grok and OSS.
- An id ending in
[1m]is marked as having a 1M context.
Claude Code lists aliases. An alias that resolves to a Claude model is named after that model; otherwise it has a name of its own:
| Alias | Shown as |
|---|---|
default | Default |
opus | Claude Opus |
sonnet | Claude Sonnet |
haiku | Claude Haiku |
fable | Claude Fable |
opusplan | Claude Opus in plan mode, Sonnet otherwise |
best | The most capable Claude model |
The effort slider
The effort control in the composer, "Reasoning effort", opens a slider. Its scale runs over seven levels:
| Level | Meaning |
|---|---|
| Minimal | Answers almost at once, with barely any reasoning |
| Low | Quick answers with light reasoning |
| Medium | Balances speed and depth |
| High | Reasons longer before answering |
| Extra high | Deep reasoning for hard problems |
| Max | The most reasoning the model allows per turn |
| Ultra | The longest reasoning, for the hardest tasks |
The slider only offers the levels the chosen model allows, as its harness reports them in its model list. Until you pick one, it shows Default, the model's own level.
- A model with no levels shows "This model has no reasoning effort levels to choose from".
- A model with a single level shows "This model thinks at a single effort level".
Snapping to the model
When you switch to a model that does not allow the current level, the effort moves to the nearest level the model allows, and the composer says so: "Max is not available for this model, so the effort is set to High". When the model allows no level at all, the effort is cleared: "Max is not available for this model".
What is remembered
- Each chat keeps its own choice. The model and effort you pick are saved with the chat in the layout, so a reopened chat continues with them.
- The last choice becomes the harness's default. Whatever you pick last for a harness is saved in
settings.json, and a new chat on that harness starts with it. A remembered effort is only applied when the new chat's model allows it.
A model or effort given explicitly, for example by a session template, comes first; then the chat's own choice; then the harness's default.
Gateway models
Gateway chats offer the models chosen in Settings › Gateway. See Gateway.