> ## Documentation Index
> Fetch the complete documentation index at: https://help.casaconect.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Choosing a model

> Which model to use for which kind of legal and property work, with a decision table, worked examples and a way to test your choice.

Choosing a model is a trade between three things: how hard the task is, how quickly you need the answer, and what you are willing to spend on it. Most of what an office asks the agent in a day is routine, and the default handles it well. A smaller share of the work is demanding enough that a stronger model pays for itself by getting it right the first time.

<Info>
  The recommendations on this page are starting points. Models differ in style as well as strength, and the best judge of a draft is the person who will sign it. See [Test your choice](#test-your-choice) for a quick way to compare.
</Info>

## Start with three questions

<AccordionGroup>
  <Accordion title="How costly is a mistake?" icon="triangle-alert">
    A status summary that misses a detail costs a minute. A deed with the wrong cadastral reference costs a great deal more. The more a mistake would cost, the stronger the model you should pick. Whichever model you use, you review the agent's edits as [tracked changes](/documents/tracked-changes) before they become part of the document.
  </Accordion>

  <Accordion title="How much does the agent have to hold in mind at once?" icon="layers">
    One short document is easy for every model. Reconciling a purchase contract against a land registry extract, a mortgage offer and three emails is not. Work that spans many sources or many steps benefits most from a frontier model.
  </Accordion>

  <Accordion title="Does where the model runs matter for this matter?" icon="map-pin">
    Claude models run on Amazon Bedrock in EU regions. ChatGPT models run on the OpenAI Platform with response storage switched off. If a client or an internal policy requires EU inference, pick a Claude model, or ask an admin to set a [model policy](/models/model-policies) for the whole organization.
  </Accordion>
</AccordionGroup>

## Model selection table

| When you need…                                               | Start with…                 | Typical work                                                                                                                                                   |
| :----------------------------------------------------------- | :-------------------------- | :------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| A quick answer at the lowest cost                            | **Auto** (ChatGPT 5.6 Luna) | Finding a document, "where does this matter stand?", filing email attachments, short summaries, renaming and tidying, simple replies                           |
| Dependable structured work                                   | **ChatGPT 5.6 Terra**       | Pulling parties, dates and amounts out of documents, filling checklists, comparing two versions of a short document, routine correspondence                    |
| Well-written drafts at a moderate price, processed in the EU | **Sonnet 5**                | Client letters and emails in Spanish or English, first drafts of standard agreements, summaries of long documents, translating between English and Spanish     |
| Strong reasoning without frontier pricing                    | **ChatGPT 5.6 Sol**         | Reviewing a contract against a checklist, spotting inconsistencies between documents, explaining the effect of a clause, preparing a list of open points       |
| Careful, nuanced drafting and review, processed in the EU    | **Opus 5**                  | Drafting and redlining deeds, contracts and powers of attorney, multi-document due diligence, long matters where the agent works through many steps on its own |
| The highest capability available                             | **ChatGPT 6 Astra**         | The hardest reasoning: reconciling conflicting sources, unusual structures, work where a second pass by a weaker model has already fallen short                |

## Two ways to settle on a model

<Tabs>
  <Tab title="Start economical">
    Leave **Auto** on and work normally. Move up only when you see a specific shortfall: a draft that reads flat, a detail missed across documents, or an instruction only half followed.

    This suits offices that use the agent mostly for intake, filing, status and correspondence, and want predictable spend.

    1. Work on Auto.
    2. When a result falls short, re-run the same request on **Sonnet 5** or **ChatGPT 5.6 Sol**.
    3. If that fixes it, use that model for that kind of task from now on.
    4. Keep Auto for everything else.
  </Tab>

  <Tab title="Start capable">
    Pick **Opus 5** or **ChatGPT 6 Astra** for substantive legal work from the first day, then step down where a cheaper model proves good enough.

    This suits offices where the agent drafts and reviews documents that go to clients, and where a lawyer's review time costs far more than the model.

    1. Do drafting and review on a frontier model.
    2. After a few weeks, try the same kinds of task on **Sonnet 5** or **ChatGPT 5.6 Sol**.
    3. Step down for the tasks where you cannot tell the difference.
  </Tab>
</Tabs>

## Worked examples

<AccordionGroup>
  <Accordion title="A client emails asking for an update on their purchase">
    **Auto.** The agent reads the matter, its linked conversations and recent activity, and drafts a reply for your approval. Nothing here needs deep reasoning, and the answer is reviewed before it is sent.
  </Accordion>

  <Accordion title="Fill a purchase deed template from the documents in the project">
    **No choice needed.** [Populate](/workflows/populate) is a workflow with its own fixed model ladder: a fast model for the bulk steps, a balanced model to identify fields and an advanced model to verify the difficult values. The model picker does not affect it.
  </Accordion>

  <Accordion title="Redline the other side's draft of a rental agreement">
    **Opus 5**, or **ChatGPT 5.6 Sol** if cost matters more. The agent must understand each clause, decide what to change and write replacement language in the right register. Edits arrive as tracked changes, so you accept or reject each one.
  </Accordion>

  <Accordion title="Check an inheritance file for missing documents">
    **ChatGPT 5.6 Terra** or **Sonnet 5.** This is checklist work across a known set of documents: thorough rather than subtle.
  </Accordion>

  <Accordion title="Three sources disagree about the surface area of a property">
    **ChatGPT 6 Astra** or **Opus 5.** Working out which source is authoritative, why they differ, and what to tell the client is the kind of reasoning frontier models are for.
  </Accordion>

  <Accordion title="Look something up on a public website">
    **Auto** or **ChatGPT 5.6 Terra.** Browsing is driven by a separate browser service; the conversation model only has to plan the visit and read the result. See [Browser sessions](/agent/browser-sessions).
  </Accordion>
</AccordionGroup>

## What each step up costs

Casa Conect charges what the provider charges, so the price gaps between models are the providers' own. As a rough guide, measured against the default model:

| Model                   | Relative cost per conversation |
| :---------------------- | :----------------------------- |
| ChatGPT 5.6 Luna (Auto) | 1×                             |
| ChatGPT 5.6 Terra       | about 10×                      |
| Sonnet 5                | about 10×                      |
| ChatGPT 5.6 Sol         | about 20×                      |
| Opus 5                  | about 25×                      |
| ChatGPT 6 Astra         | about 45×                      |

These multiples sound large, but the base is small. A long drafting conversation on a frontier model usually costs less than a few minutes of a lawyer's time. See [AI credits](/admin/ai-credits) for list prices and for how to set budgets.

<Tip>
  Very long conversations cost more per message, because the model re-reads the history on every turn. When you change subject, start a new conversation. It is cheaper and the agent stays focused.
</Tip>

## Test your choice

You do not need a formal evaluation. Take a task you know well and run it twice.

<Steps>
  <Step title="Pick a real task with a known good answer">
    Use a document you have already reviewed by hand, so you can tell at a glance what the agent gets right and wrong.
  </Step>

  <Step title="Run it on two models">
    Send the same request in two separate conversations, one per model. Use a copy of the document if the task edits it.
  </Step>

  <Step title="Compare what matters to you">
    Look at accuracy first, then completeness, then tone. Check the cost of each conversation under **Settings → Access & billing → Usage**.
  </Step>

  <Step title="Write down the result">
    If a cheaper model is good enough for this kind of task, tell your team. An admin can also narrow the list under [Model policies](/models/model-policies) so the choice is simpler.
  </Step>
</Steps>

## Getting more from any model

The model matters less than the instruction. Before moving up a tier, check that you have given the agent what it needs.

* **Attach or mention the documents** you want used rather than describing them.
* **Say what good looks like**: who the reader is, the language, the length and the format.
* **Use a [skill](/agent/skills)** for work you repeat, so the method is written down once.
* **Add [guidance](/agent/guidance)** for house style and standing preferences, so you do not repeat them in every message.

<CardGroup cols={2}>
  <Card title="Working with the agent" icon="message-square" href="/agent/best-practices">
    Habits that improve results on every model.
  </Card>

  <Card title="Model policies" icon="toggle-right" href="/models/model-policies">
    Restrict the list for your organization.
  </Card>
</CardGroup>