> ## Documentation Index > Fetch the complete documentation index at: https://docs.textql.com/llms.txt > Use this file to discover all available pages before exploring further. # Model Management > Configure which AI models are available in your organization, set defaults, and control per-role access TextQL supports a range of AI models from Anthropic, OpenAI, and other providers. Organization administrators can control which models are available, set defaults, and decide whether users can switch models themselves. Model settings are found under **Settings → Models**, which has three tabs: **Model Catalog**, **Role Access**, and **Analytics**. *** ## Available Models TextQL's model catalog includes models across several providers: ### Anthropic (Claude) | Model | Best For | Context Window | | - | - | - | | Claude Fable 5.2 | Latest Fable. Demanding reasoning and long-horizon agentic work. Lower cost than Fable 5.1. Thinking is always on and cannot be disabled. Requires the "Allow LLM Data Retention" security setting to be enabled, not available under zero data retention. | 1M | | Claude Fable 5.1 | Deprecated, retires December 15, 2026 (moves to Claude Fable 5.2). Previous Fable release. Demanding reasoning and long-horizon agentic work. Requires the "Allow LLM Data Retention" security setting to be enabled, not available under zero data retention. | 1M\* | | Claude Fable 5 | Deprecated, retires December 15, 2026 (moves to Claude Fable 5.2). Previous Fable release. Demanding reasoning and long-horizon agentic work. Requires the "Allow LLM Data Retention" security setting to be enabled, not available under zero data retention. | 1M\* | | Claude Opus 5.5 | Latest Opus. Complex agentic coding, deep reasoning, and long-horizon enterprise work. Thinking is always on and cannot be disabled. Lower cost than Opus 5. Supports [Fast Mode](https://docs.textql.com/core/admin/pricing.md#inference) (2× ACU cost). | 1M\* | | Claude Opus 5 | Deprecated, retires December 15, 2026 (moves to Claude Opus 5.5). Previous Opus release. Complex agentic coding, deep reasoning, and long-horizon enterprise work. Higher cost than Sonnet. Supports [Fast Mode](https://docs.textql.com/core/admin/pricing.md#inference) (2× ACU cost). | 1M\* | | Claude Opus 4.8 | Deprecated, retires December 15, 2026 (moves to Claude Opus 5.5). Most capable Anthropic model. Complex analysis, advanced reasoning, demanding agentic tasks. Higher cost than Sonnet. Supports [Fast Mode](https://docs.textql.com/core/admin/pricing.md#inference) (2× ACU cost). | 1M\* | | Claude Opus 4.7 | Complex analysis, advanced reasoning, demanding agentic tasks. Higher cost than Sonnet. | 1M\* | | Claude Sonnet 5.5 | Latest Sonnet. Near Opus 4.8 quality for most tasks without the Opus cost. Adaptive thinking is always on and cannot be disabled. On Amazon Bedrock, available only when global cross-region inference is enabled. | 1M\* | | Claude Sonnet 5 | Deprecated, retires December 15, 2026 (moves to Claude Sonnet 5.5). Previous Sonnet release. Near Opus 4.8 quality for most tasks, SQL generation, dashboards, agentic workflows, at a fraction of the cost. | 1M\* | | Claude Sonnet 4.6 | Most tasks — SQL generation, dashboards, agentic workflows. | 1M\* | | Claude Haiku 5.5 | Latest Haiku. Fastest and lowest cost, for Q\&A, lookups, and lighter agentic work. Prompts over 100K tokens bill at a higher rate. On Amazon Bedrock, available only when global cross-region inference is enabled. | 1M† | | Claude Haiku 4.5 | Deprecated, retires December 15, 2026 (moves to Claude Haiku 5.5). Simple Q\&A and quick lookups. Fastest and lowest cost. | 200K | \* 1M token context window available in beta. Standard pricing applies up to 200K input tokens; tokens beyond 200K are billed at a higher rate. See [Pricing](https://docs.textql.com/core/admin/pricing.md#inference) for details. † Standard pricing applies to prompts up to 100K input tokens. A prompt over 100K tokens bills every token in the request at 5× the standard rate. See [Pricing](https://docs.textql.com/core/admin/pricing.md#inference) for details. ### OpenAI (GPT) | Model | Best For | Context Window | | - | - | - | | GPT-6 Astra | OpenAI's newest flagship, for the most demanding reasoning and agentic work. | 1.05M | | GPT-6.1 Sol | OpenAI's newest Sol. An upgrade over GPT-6 Sol at the same price, with cheaper cached input. | 1.05M | | GPT-6.1 Terra | Balanced GPT-6.1 tier. Strong reasoning and coding at half the price of GPT-6.1 Sol. | 1.05M | | GPT-6 Sol | Deprecated, retires December 15, 2026 (moves to GPT-6.1 Sol). Balanced GPT-6 model for coding, reasoning, and agentic workflows. | 1.05M | | GPT-6 Luna | Efficient model for focused, high-volume tasks. | 1.05M | | GPT-5.6 Sol | Deprecated, retires December 15, 2026 (moves to GPT-6.1 Sol). Advanced coding, research, and the most demanding agentic workflows. | 1M | | GPT-5.6 Terra | Deprecated, retires December 15, 2026 (moves to GPT-6.1 Terra). GPT-5.5-level performance at roughly half the cost. Good OpenAI default for most tasks. | 1M | | GPT-5.6 Luna | Deprecated, retires December 15, 2026 (moves to GPT-6 Luna). High-throughput, routine tasks. Fastest and most cost-efficient GPT-5.6 tier. | 1M | | GPT-5.5 | Deprecated, retired November 2, 2026 (moved to GPT-6 Sol, which moves to GPT-6.1 Sol on December 15, 2026). Complex reasoning, dashboards, and agentic workflows. | 1M\* | | GPT-5.4 | Deprecated, retired November 2, 2026 (moved to GPT-6 Sol, which moves to GPT-6.1 Sol on December 15, 2026). Complex reasoning, dashboards, data analysis. | 1M\* | | GPT-5.4 Mini | Deprecated, retires November 2, 2026 (moves to GPT-6 Luna). Simple, well-defined tasks. | 1M\* | ### Google (Gemini) | Model | Best For | Context Window | | - | - | - | | Gemini 3.1 Pro | Google's most advanced reasoning model. Complex reasoning and agentic workflows. | 1M\* | | Gemini 3.8 Flash | Google's newest Flash model. Frontier coding and agentic performance at Flash latency and cost. Currently on introductory pricing through Dec 31, 2026. | 1M | | Gemini 3.5 Flash | Deprecated, retires December 15, 2026 (moves to Gemini 3.8 Flash). Frontier coding and agentic performance at Flash latency and cost. | 1M\* | ### Meta | Model | Best For | Context Window | | - | - | - | | Muse Spark 1.2 | Meta's latest frontier model. Tool use and multi-step work with always-on reasoning. US-only preview. | 1M | | Muse Spark 1.1 | Deprecated, retires December 15, 2026 (moves to Muse Spark 1.2). Meta's first frontier model. Tool use and multi-step work with always-on reasoning. US-only preview. | 1M | ### DeepSeek (via Fireworks) | Model | Best For | Context Window | | - | - | - | | DeepSeek V4.1 Flash | DeepSeek's newest reasoning model with image input and tool calling, for fast and very low-cost high-throughput work. | 1M | | DeepSeek V4 Flash 0731 | DeepSeek's fast, very low-cost model. Good for high-throughput and routine analytical tasks. | 1M | | DeepSeek V4 Flash Vision Exp | DeepSeek's low-cost V4 Flash with image input. Good for reading charts, screenshots, and scanned documents. | 1M | ### Moonshot (via Fireworks) | Model | Best For | Context Window | | - | - | - | | Kimi K2.7 | Kimi frontier model. Long-horizon coding, agentic execution, long-context reasoning. | 256K | | Kimi K3 | Latest Kimi frontier model. Multimodal; repository-scale coding, debugging, long-horizon agentic work. | 1M | ### Zhipu / GLM (via Fireworks) | Model | Best For | Context Window | | - | - | - | | GLM 5.3 | Frontier GLM reasoning model. Complex agentic workflows and long-context analysis. | 1M | | GLM 5.3 Flash | Fast, cost-efficient multimodal analysis and high-throughput agentic work. | 1M | | GLM 5.2 | Frontier GLM model. Complex reasoning and very-long-document analysis. | 1M | ### Qwen (via Fireworks) | Model | Best For | Context Window | | - | - | - | | Qwen3.7 Plus | Cost-effective coding and general analytical tasks with a large context window. | 256K | For ACU pricing on each model, see the [Pricing page](https://docs.textql.com/core/admin/pricing.md#inference). *** ## Model Catalog The **Model Catalog** tab is where administrators enable or disable models for the organization and set the organization-wide default. ![Model Catalog](https://mintcdn.com/textql/ZhijkTai56smTOBc/images/admin/models/model-catalog.png?fit=max\&auto=format\&n=ZhijkTai56smTOBc\&q=85\&s=a14f7cacbda74d29891d8a31604d048a) Model Catalog tab in Settings → Models ### Enabling and Disabling Models By default, all models in the catalog are available to your organization. You can disable any model to prevent it from appearing in the model picker for all users. A disabled model cannot be selected by any user regardless of their role settings. If a role has a disabled model set as its default, the system default takes over. Use **Enable all models** above the table to turn every model back on in one step. ### Automatically Enabling New Models The **Automatically enable new models** switch above the catalog decides what happens when a model is added to the catalog later. When it is on, new models are enabled for the organization even if you have disabled others; only the models you turned off stay off. When it is off, the catalog is pinned to the models enabled today and new models start disabled until you turn them on. New organizations start with this switch on and every model enabled. ### Setting the Organization Default The organization default is the model Ana uses when no specific model has been selected for a conversation. It appears as **"Org Default"** in role settings and is distinct from the **System Default**. * **System Default**: TextQL's platform-level default. On TextQL's hosted service it is currently Claude Sonnet 5.5 for Anthropic organizations and GPT-6.1 Sol for OpenAI organizations, and it updates automatically as the platform evolves. On a self-hosted deployment a new default takes effect only after a platform admin accepts it (see [Model Deprecations](#model-deprecations)); until then Anthropic organizations keep Claude Sonnet 5 and OpenAI organizations keep GPT-5.6 Sol. Setting `MODEL_DEPRECATIONS_ENABLED` overrides acceptance: `true` applies every change without it, and `false` applies none even when accepted. * **Org Default**: An explicit model your organization has pinned. Overrides the System Default for all roles that inherit from the org. If no org default is set, all roles fall back to the System Default. *** ## Role Access The **Role Access** tab gives fine-grained control over which models each role can use, what their default is, and whether users in that role can switch models mid-conversation. ![Role Access](https://mintcdn.com/textql/ZhijkTai56smTOBc/images/admin/models/role-access.png?fit=max\&auto=format\&n=ZhijkTai56smTOBc\&q=85\&s=dc2485a0c5675ccb9bfd81f60e8c12c5) Role Access tab showing per-role model configuration Each role has three configurable settings: ### Default Model The model Ana starts a new conversation with for users in this role. Options are: * **Org Default** — Inherits whatever the organization default is set to (recommended for most roles) * **System Default** — Pins to TextQL's platform default, auto-updating with the platform * **Specific model** — Pins to an explicit model regardless of org or system defaults ### Allowed Models The subset of org-enabled models that users in this role can see in the model picker. Defaults to all enabled models. Restricting this is useful if you want certain roles to only access lower-cost or specific provider models. The allowed models list has **Select all** and **Clear all** controls. **Clear all** keeps the role's default model, since a role always keeps access to its default. **Automatically include new models** controls models added later, whether they are new to the catalog or newly enabled for your organization. When it is on, those models are allowed for the role without editing its list; only the models you unchecked stay excluded. When it is off, the role keeps exactly the models checked today. > **Note:** A role's default model is always included in its allowed list. You cannot remove a model from the allowed list if it is set as that role's default — change the role default first. When the organization has an explicit default model, every role that keeps **Org Default** can use it even if it is not checked in the role's allowed list, so changing the org default reaches those roles without editing each list. If the organization follows the System Default instead, the role's allowed list still applies and users start on the first allowed model when the platform default is not in it. When a user holds several roles, the most permissive role wins: the user can switch models if any role allows it, and can use any model that any of their roles allows. A role with all models allowed grants access to every enabled model regardless of the other roles' restrictions. ### Can Switch Use **Let every role switch models** above the table to turn on switching for every role at once, then save. The role editor (**Settings → Roles → Edit**) has the same switch, and if model switching is turned off for the whole workspace it shows a **Turn on** button so you can fix both from one place. Controls whether users in this role see the model picker in the chat interface at all. When disabled, users in this role cannot change the model — they always use the role's default. | Can Switch | Behavior | | - | - | | On | Users see the model picker and can choose any model in their allowed list | | Off | Model picker is hidden; the role's default model is used silently | ### Configuring via API These three settings are fields on the role, set from this tab or programmatically with the role update API (`RBACService/UpdateRole`). They are **not** [RBAC permissions](https://docs.textql.com/core/admin/rbac.md) and cannot be granted through a role's permission list — `allow_model_switching` is a UI label, not a permission string. | UI setting | Role field | | - | - | | Can Switch | `allow_model_choice` | | Default Model | `default_model_id` | | Allowed Models | `allowed_model_ids` | | Automatically include new models | `include_new_models` | `default_model_id` and `allowed_model_ids` take `LlmModel` enum numbers. An empty `allowed_model_ids` is treated as "no change"; send `clear_allowed_model_ids: true` to reset it back to all models. *** ## Model Picker (User-Facing) When model switching is enabled for a role, users see a model picker in the top right of the chat interface. It shows: * **Default** — The system/org/role default, auto-updating with the platform * All models the role is allowed to access, grouped with provider icons Hovering over a model for \~1 second expands its description. The currently selected model is highlighted with a checkmark. Models approaching deprecation show a warning icon with the deprecation date. ![Model Picker](https://mintcdn.com/textql/ZhijkTai56smTOBc/images/admin/models/model-picker.png?fit=max\&auto=format\&n=ZhijkTai56smTOBc\&q=85\&s=6e4451a55a3873c450610d9c749a36ec) Model picker in the chat interface A model selection is scoped to that conversation — switching models in one chat does not affect other chats. *** ## Thinking Levels For models that support it, members choose how long the model thinks before answering. The level is set from the gauge button next to the model picker in the composer, or from the **Thinking** slider at the bottom of the model picker. It can be changed mid-thread and applies from the next message. | Level | When to use | | - | - | | Off | The fastest replies, with no thinking before the answer | | Low | Quick answers and simple lookups | | Medium | Everyday analysis | | High | Multi-step analysis where accuracy matters | | Extra high | Long agentic work, such as building dashboards or apps | | Max | The hardest problems, when correctness matters more than speed or cost | Each model shows only the levels it accepts. A thread with no level chosen sends nothing, so the model runs at its provider's own default, shown in the last column. | Models | Levels | Default | | - | - | - | | Claude Haiku 5.5 | Off to Max | Medium | | Claude Opus 5.5 | Low to Max | Medium | | Claude Opus 4.8, Sonnet 5.5, Sonnet 5 | Off to Max | High | | Claude Opus 5, Fable 5.2, Fable 5.1, Fable 5 | Low to Max | High | | GPT-6.1 Terra, GPT-6 Sol, GPT-6 Luna, GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna | Off to Max | Medium | | GPT-6.1 Sol | Low to Max | Medium | | GPT-6 Astra | Low to Extra high | Medium | | Gemini 3.1 Pro | Low to High | High | | Gemini 3.5 Flash | Off to High | Medium | | Gemini 3.8 Flash | Low to High | Medium | | Kimi K3, DeepSeek V4.1 Flash (via Fireworks) | Off to High | Medium | | Other DeepSeek V4, GLM 5.x, and Kimi models (via Fireworks) | Low to High | Medium | Other models don't show the selector. Claude levels apply on the native Anthropic API, Amazon Bedrock, and Google Vertex. At every level but Off, Claude thinks with a visible summary. GPT levels apply on the native OpenAI API; Azure and other OpenAI-compatible gateways run GPT models at their default. **Off** is offered only where the provider can stop the model thinking: Claude turns thinking off (Sonnet 5.5 keeps only its short notes between tool calls), GPT and the Fireworks models run with reasoning effort `none`, and Gemini 3.5 Flash runs at its minimal level. A model without Off, such as Claude Opus 5.5 or GLM 5.3, runs at Low instead. When a member picks a level, it becomes their starting level for new threads on that model, in that browser. Admins set the defaults for everyone else in two places: * **Org default:** **Settings > Capabilities > Thinking Levels > Default thinking level** applies to every model. * **Per-model default:** the **Default thinking** column under **Settings > Models > Model Catalog** overrides the org default for one model. Choose **Use org default** to remove the override. A new thread starts at the first of these that is set: the member's last pick for that model, the model's per-model default, the org default, then the provider's default. This applies to threads created through the API too. A model that tops out below the chosen level uses its highest level, and a model without Off uses Low. The org default stays a fallback, so models added later pick it up without any change. The **Default thinking** column shows each model's starting level and where it comes from, for example **Override: Low**, **Org default: High**, or **Provider default: Medium**, and the composer shows the same source under the thinking slider. ### Agents and playbooks An agent runs at its own thinking level, set in the agent editor under **Advanced settings**, next to its model. The editor shows only the levels the agent's model offers. An agent with no level runs at the provider's default. Org and per-model defaults never apply to agents, so changing them never changes what an agent costs to run. The agent's level applies to every run, whether it was scheduled, run manually, mentioned in Slack or Teams, chatted with, or called as a subagent. An agent that follows the workspace default model is held to that model's levels when it runs. Agents defined in config files set the level with `thinking_level` next to `model`; a level the model doesn't offer is rejected when the file is saved. Playbooks always run at the provider's default. When a model is retired, each agent on it that isn't defined in a config file moves to the successor model, and its level drops to the nearest level the successor offers, for example **Max** on a retired model becomes **Extra high** on GPT-6 Astra. A per-model default for the retired model carries to the successor only when the successor has no default of its own. The retirement email lists each agent with its level before and after and says whether the per-model default carried over; the in-app notification gives the number of agents moved and the per-model default change. An agent defined in a config file is updated by a change to its file, which rewrites `model` and `thinking_level` together. **Cost.** Pricing per token doesn't change. Thinking tokens bill as output tokens, so higher levels cost more per message. Admins can turn the selector off for the whole organization with the **Thinking Levels** capability. When it's off, every thread and every agent run uses the model's default level. *** ## Model Deprecations When a model is scheduled for deprecation, a warning appears in the model picker next to that model's name, showing the deprecation date. Administrators should update any role defaults that point to a deprecating model before the deprecation date to avoid disruption. After deprecation, requests to a deprecated model are automatically routed to the nearest equivalent current model. On a self-hosted deployment (single-tenant or VPC), a model deprecation does nothing until a platform admin accepts it under **Settings > Platform admin > Model management**. That page lists the deprecations that ship with your release, shows whether your model endpoint serves each model and its successor and how much of your workspace uses it, and lets you accept a deprecation as shipped, choose a different successor your endpoint serves, move its date, or skip it. **Test successor** makes one real call through your endpoint before you accept. An accepted deprecation sends the same warnings and emails and moves playbooks, defaults and threads on its date. Setting `compute.modelDeprecationsEnabled` in your Helm values turns every deprecation on or off at once instead. The same page decides your deployment's **system defaults**: the model each tier uses when nothing more specific is set, such as a new thread with no model picked or a background summary. When a release moves a default, for example from Claude Sonnet 5 to Claude Sonnet 5.5, a self-hosted deployment keeps its current default until a platform admin accepts the change, either as shipped or with another model from the same provider that your endpoint serves. **Test model** makes one real call before you accept. A deprecation of a model that is still a tier's default can only be accepted together with the change that moves the tier off it. `SYSTEM_DEFAULT_MODEL` still overrides the default tier. *** ## Frequently Asked Questions **Can different roles use different providers (Anthropic vs OpenAI)?** Yes. You can configure one role to default to Claude Sonnet and another to GPT-5.2 by setting explicit role defaults and restricting allowed models per role. **What happens if a user's role default model is disabled at the org level?** The system falls back to the org default, then the system default. Disabled models are never served to users even if a role still references them. **Does model selection affect Playbooks?** Playbooks use the organization's default model unless a specific model has been configured for that playbook. Role-level model restrictions do not apply to Playbook runs. **Can I restrict users to only Anthropic models?** Yes — set the allowed models for a role to only include Claude models. Users in that role will not see OpenAI or other provider options in the picker. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.