Reasoning effort and visible reasoning
Next to the AI selector, the assistant and Edit with AI offer a Reasoning effort: how hard the model thinks before it answers. More effort means deeper reasoning, slower answers and more tokens.
The levels
| Level | Shown as |
|---|---|
low | Low |
medium | Medium |
high | High |
xhigh | Extra high |
max | Max |
You choose it per question in the assistant, and per request (and per follow-up) in Edit with AI. When you choose none:
- the assistant answers at medium;
- a new Edit with AI request runs at high;
- a follow-up runs at the effort of the message before it.
The effort used is shown next to the tokens of each answer and recorded with the usage.
Reasoning counts towards what a model may write in one turn, so that ceiling grows with the effort: for
the assistant 16K tokens up to medium, 32K at high and 64K at xhigh and max; for the code agent
32K up to high and 64K above.
Which providers take it
| Provider | Levels offered |
|---|---|
| Claude (Anthropic, Azure AI Foundry, AWS Bedrock) | All five |
| GPT on Azure OpenAI with Reasoning model on | Low, medium, high; xhigh and max are sent as high |
| GPT on Azure OpenAI without reasoning, Gemini | None: the choice is hidden |
The levels on Claude need Claude Opus 4.7, Sonnet 5 or later. The default model, claude-opus-5-5,
is one of them. On Foundry the deployment, and on Bedrock the model id, must be such a model: an older
one rejects the requests.
Claude's reasoning
Claude returns a summary of its reasoning while it works:
- in the assistant it streams live under Reasoning, above the answer, and stays in the saved conversation;
- in Edit with AI it appears in the agent's progress log, the first 1,000 characters of each turn.
Secrets are masked in what is kept, as in the answers. The summary changes what you see, not what is billed. The other providers show no reasoning.
Cap the effort
An AI can have a Maximum effort (the system AI takes it from AI_MAX_EFFORT): the console offers no
level above it, and a request asking for more is served at the cap, which is what gets stored. Without a
cap every level the provider takes is offered.
On an installation open to sign-ups the cap bounds the cost of a single question: at max each model
turn may write up to 64K tokens.
See AIs of an organization to set the cap.