> ## 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. # Models overview > The AI models available in Casa Conect, who makes them, where they run, and how the Auto setting works. The Casa Conect agent is not tied to a single AI model. Each conversation runs on a model you choose from the composer, or on the default that **Auto** selects for you. This page lists what is available and how the choice is applied. For advice on which model suits which kind of work, see [Choosing a model](/models/choosing-a-model). ## Available models | Model | Made by | Runs on | Position | | :-------------------- | :-------- | :------------------ | :------------------------------------------------------------------ | | **ChatGPT 6 Astra** | OpenAI | OpenAI Platform | Frontier. The most capable and the most expensive model on offer. | | **Opus 5** | Anthropic | Amazon Bedrock (EU) | Frontier. Deep reasoning and careful long-form drafting. | | **ChatGPT 5.6 Sol** | OpenAI | OpenAI Platform | Advanced. Strong reasoning at a lower price than the frontier pair. | | **Sonnet 5** | Anthropic | Amazon Bedrock (EU) | Balanced. Good drafting quality at a moderate price. | | **ChatGPT 5.6 Terra** | OpenAI | OpenAI Platform | Balanced. Reliable structured work and extraction. | | **ChatGPT 5.6 Luna** | OpenAI | OpenAI Platform | Fast and economical. The model behind Auto. | The same list appears in the dashboard composer and in the [Word add-in](/word-add-in/overview). The model line-up changes as providers release and retire models. If a model you had selected is retired, your composer falls back to Auto rather than failing. ## The model picker Open the model menu in the composer toolbar to see the list. The first row is the **Auto** switch; below it are the individual models, each with its maker's mark. Auto is on by default. It sends your conversation to Casa Conect's default model, chosen for speed and low cost on routine requests. Today that is ChatGPT 5.6 Luna. Turn Auto off and choose a model. Your choice is remembered in that browser for future conversations until you change it. You can change model between messages in the same conversation. The next message runs on the new model and the conversation history carries over. Auto is a fixed default, not a router that inspects each request. If you are about to ask for demanding work, such as redlining a long contract or reconciling several documents, pick a stronger model yourself before you send the message. ## Where models run Opus 5 and Sonnet 5 run on **Amazon Bedrock** through its EU inference profile, so requests are processed in EU regions. Astra, Sol, Terra and Luna run on the **OpenAI Platform**. Casa Conect sends every request with response storage switched off, so OpenAI does not retain the conversation as a stored response. Read [AI data handling](/security/ai-data-handling) for the full picture, including the supporting models used for search, document parsing and dictation. ## Models that work behind the scenes Some work never reaches the model picker because Casa Conect assigns a fixed model to it: | Work | What runs it | | :--------------------------------------------- | :-------------------------------------------------------------------------------------------------------------------------------------- | | Naming a conversation | The fast default model | | Applying a matter template | The fast default model | | Shared space summaries | The fast default model | | [Populate](/workflows/populate) | A three-step ladder: a fast model for bulk steps, a balanced model to identify fields, and an advanced model to verify difficult values | | [Anonymize](/workflows/anonymize) | A Claude model to identify sensitive values, and the fast default model for the remaining steps | | Helpers the agent starts during a conversation | The same model as the conversation that started them | | Searching your documents | Dedicated embedding and reranking models. See [Search](/documents/search). | | Voice dictation | A speech-to-text model | [Model policies](/models/model-policies) govern the model a member picks for an agent conversation. Background helpers use the fixed models above regardless of policy. ## Limits that apply to every model These limits protect you from runaway conversations. They are the same whichever model you pick. | Limit | Value | | :------------------------------------------------------------------ | :--------- | | Steps the agent may take in one turn, when you are chatting with it | 50 | | Steps in an unattended turn started by a verified member's email | 30 | | Steps in a turn started by an external sender's email | 10 | | Maximum length of a turn | 20 minutes | | Maximum length of a single step | 4 minutes | The final step of a turn is always reserved for a written answer, so a turn that reaches its budget still ends with an explanation of where it got to. ## Next steps Match the model to the work: a decision table and worked examples. Let admins decide which models the team can use. How model usage is metered and what each model costs. What is sent to model providers, and what is not.