Some models answer. Some plan before they answer.

In Part 5 you saw the collapsible block that sometimes appears above an answer. It is the visible edge of a real difference in how a model works.

• A general model answers directly. Good for a short lookup, a simple rewrite, one tool call.
• A reasoning model spends extra computation planning or checking first. That can help with an automation spanning several rooms, devices and time periods.

Two consequences follow. Usage: some providers bill reasoning as output, others as a separate category — accounting depends on model and route. Latency: it may take longer to answer, with no universal delay and no guarantee.

The one setting that trips people up

The Add Model dialog has a field called thinkingFormat. It tells Pi Agent how a route represents reasoning data; it does not make a model reason.

Deciding which jobs deserve the extra thinking

One rule covers most cases: use a general model when the result is easy to verify and the downside small. Use reasoning when several constraints must be balanced, or an error would be expensive.

flowchart TD
  A["A task"] --> B{"Can you check the
answer in a minute?"}
  B -->|"yes"| C["General model"]
  B -->|"no"| D{"Would being wrong
cost money or comfort?"}
  D -->|"no"| C
  D -->|"yes"| E{"Several constraints
at once?"}
  E -->|"no"| C
  E -->|"yes"| F["Reasoning model,
then verify anyway"]

Against real tasks:

Where model and route support it, you get a dial rather than an on/off switch.

Reasoning levelHow much thinking to allow before the answer. Medium is selected here.

• auto hands the decision to the route; off asks for none at all. medium carries the tick here.
• high, xhigh, max buy more thinking, more waiting and more usage in one move.
• These controls are model-specific. A route may not honor every level, so check the answer rather than assume the dial did something.

To genuinely cut reasoning usage, choose a documented non-reasoning model, or use a supported control for that exact model. Do not invent parameters.

Changing model without starting over

The selector sits in the toolbar below the composer — the text box. It is a button with a small arrow showing the current model name. Its position moves around, because the toolbar order is configurable, so look for the name, not a fixed spot.

It works by mouse or keyboard. The search field at the top filters the list by any part of a model name — typing sol leaves GPT-5.6 Sol — and Enter selects the match. Handy once several providers fill it up.

The selectorEvery model under every provider you added.

• The tick beside GPT-5.6 Sol marks the model that answers your next message.
• Every name here is a GPT model, GPT-5.3 Codex Spark down to GPT-6 Astra. A provider missing entirely was never saved in the Models panel.
• The model button under the message box repeats the current name, GPT-5.6 Sol.

Two details are worth getting straight, because almost every later surprise comes from one.

• It takes effect immediately, but nothing is rerun. The new model answers your next message. Earlier messages are not resubmitted and earlier answers do not disappear, so nothing already paid for is charged again.
• The new model receives the whole conversation. On its first message it reads the entire Session, first line to last, as input tokens. That is how it follows the topic — and it is the cost below.

The pattern most people end up using every day

This assumes two models are available: one fast and cheap, one stronger. The example uses glm-4-flash and claude-sonnet-4-5; any pair of that shape works.

1. Step 1

   Start a new Session and pick the cheap model

   Select + New session at the top left. Open the model button below the composer, find the GLM group and choose glm-4-flash. The label changes.

2. Step 2

   Ask for the draft

   Send: “Write a Home Assistant automation that turns off every light in the living room at 11:30 p.m. If someone is still in the living room at that time, wait 10 minutes before turning them off.” Expect YAML in roughly 20-40 seconds.

3. Step 3

   Switch to the stronger model — without clearing anything

   Stay in the same conversation; do not start a new Session. Open the selector and choose claude-sonnet-4-5 under Anthropic. The label changes.

4. Step 4

   Ask it to review what is already on screen

   Send: “Please review the automation above. Does it have any logic errors or common HA pitfalls, such as the device_id trap, the wrong mode, or failure to detect a nonexistent entity_id?” It reads the whole conversation, YAML included.

5. Step 5

   Switch back and let the cheap model apply the notes

   Return to glm-4-flash and ask it to rewrite the YAML using the feedback above. Applying a list of corrections does not need the expensive model.

The expensive model handles one message in three: the one where judgment is the product. A review typically turns up device_id where entity_id belongs, a mode that should be restart not the default single, or triggers and conditions in the older singular form. The saving over using the strong model throughout is commonly observed at 50-80%; an earlier passage in the same source says 60-80%. Both depend on prompt length and model choice, so treat either as a rough shape rather than a quote.

Every switch makes someone read the transcript again

This is what surprises people when the first invoice arrives.

As numbers. A conversation has built up 5,000 tokens, roughly 15-20 exchanges, and you switch four times:

Stay on one model and input still grows with each turn, but none of this rereading happens. That gap is the whole cost of switching. And you can watch it, because the workspace prints the meter under every answer:

Where the reread shows upA one-sentence question, and 1,094 tokens of input. That is the figure a switch resets.Live workspace

• 1,094 in against 53 out. Input is already the large half on a short exchange, and only grows.
• Watch this line after a switch. Input is the number that jumps, not output.
• $0.0071 is nothing once. Multiplied by a long Session and four switches, it stops being nothing.

Three ways to hand work over, and one that fails quietly

The selector is not your only option. Three actions exist and they do different jobs. Say Flash has just recommended a 1 ton air conditioner for your living room:

Three patterns are worth switching for: draft-then-review, zooming in on one hard point before returning to the cheap model, and A/B comparison — which is really a fork. A switch fitting none of them is probably costing you a reread for nothing.

The switch that fails silently

Claude represents reasoning as native structured blocks, not plain text. OpenAI-compatible providers — GLM, DeepSeek, Groq, OpenRouter — do not use that structure.

So those blocks cannot enter the new provider's messages array in their original form. Depending on the version they are discarded or flattened into plain text; both have happened. If your next message says “continue with the steps you just considered,” the new model may not have them.

Where to go from here