Choosing an AI model

Strata runs every prompt through one of several frontier AI models. The exact set you can pick from is a curated Model Catalog that Kronisys maintains and your organization enables — this page covers how the catalog works, how to choose, and how to switch.

Where the model list comes from

Strata is built on a DB-driven Model Catalog. Kronisys (the platform owner) curates a central catalog of models across many providers — OpenAI, Anthropic, xAI, DeepSeek, Mistral AI, Cohere, and Microsoft — all served through Microsoft Azure AI Foundry. Your administrator then enables the subset your organization wants.

What this means for you:

• The models in your picker are the ones you may use: what your organization has enabled from the catalog, narrowed by your role and by any limit an administrator has set for you. Two organizations, or two people in one, can see different lists.
• The catalog can grow or change without an app update. A model Kronisys adds and your admin enables simply appears; a retired one disappears.
• The in-app picker is always the live source of truth for which models you can use right now.

Note

Representative chat models in the catalog include GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, Claude Sonnet 5, the GPT-5.6 family (Sol, Terra, and Luna), GPT-5.4, GPT-5.4 Mini, Grok 4.6, DeepSeek V4 Pro, and Mistral Medium 3.5 (in preview on Azure AI Foundry) — alongside earlier Claude models and other Grok, DeepSeek, Mistral, Cohere, and Phi models. Your organization decides which of these are turned on, so your list may be shorter. In an organization whose admin has chosen its models, a model Kronisys adds is not switched on automatically: an admin enables it once it is deployed in your organization's own Azure AI Foundry.

Switching the model

1. In the chat composer, click the model name in the toolbar to open the model picker.
2. Click the model you want. Current models are listed first. Older models are under the Legacy models row at the bottom: point at it (or click it) and they open in a list beside the dropdown. A legacy model with a retirement date shows it under its name, for example Retires Oct 19 (the year is added when it isn't this year). On a phone, where there is no room beside it, the list opens over the model list instead, with a row at the top that takes you back.

The pick takes effect on your next message. It is saved as this conversation's model and becomes the default for new chats — there's no separate "default model" setting to manage. See Settings for the rest of your account preferences.

Auto: let Strata pick

Auto is the first choice in the model picker. With Auto on, Strata picks the model for each message from the ones you may use:

• Everyday messages (short questions, rewrites, replies, quick lookups) go to a fast model: the first of GPT-5.4 Mini, Claude Haiku 4.5, GPT-6 Luna, GPT-5.6 Luna, Grok 4.1 Fast and DeepSeek V4 Flash that is open to you.
• Demanding messages go to a stronger model: the first of Claude Sonnet 5.5, GPT-6 Sol, Claude Sonnet 5, GPT-5.4, GPT-5.6 Terra, Claude Sonnet 4.6, GPT-5.6 Sol, Grok 4.6, Claude Sonnet 4.5 and DeepSeek V4 Pro that is open to you.

A message counts as demanding when it:

• has files or images attached, or pasted text;
• is about a document (a workbook, a PDF or report, a deck, a Word document, a memo, a dashboard, a proposal or contract);
• involves data work (SQL, tables, forecasts, trends), or asks something like "how many orders by region?" while a database is connected;
• asks for a chart or a calculation (a graph, a what-if, a break-even, an estimate);
• asks for analysis, a comparison, a plan, a review or research;
• involves code;
• works with your email or inbox, your calendar or meetings, OneDrive or SharePoint, or Jira tickets;
• mentions something with @;
• is long (over 400 characters), or comes late in a long conversation;
• or follows up on any of those. "And for Q2?" after a revenue question, or "yes" after Strata offers to build a deck, stays with the stronger model. A closing "thanks" goes to the fast model.

Seeing which model answered

Every answer Auto wrote says which model it picked, on a line under the answer, for example Auto · Claude Sonnet 5.5. The line appears as soon as Auto picks, before the answer starts, and stays with the answer when you reopen the chat. Point at it to see why Auto picked that model ("a stronger model, for data work and a chart or a calculation"); screen readers read the same sentence.

Things to know

• Auto only picks a model you may use. Your organization, your role and any limit an administrator has set for you narrow its choices exactly as they narrow the picker. If none of the models above is open to you, it uses another model you may use.
• Auto leaves out the top-tier models such as Claude Opus 5.5 and GPT-6 Astra, which cost several times as much, unless no other model is open to you. For the hardest problems, choose one by name.
• With images attached, Auto picks a model that can see them when one is open to you (DeepSeek models read text only).
• If your organization runs its own Azure AI Foundry, Auto leaves out a model that is not deployed there. It checks when you choose Auto and refreshes that check at most every 20 minutes, and a message that finds a model missing makes Auto skip that model until the next check.
• Your Effort level and the Thinking switch apply to whichever model Auto picks, and a usage allowance on a model applies when Auto picks it, just as if you had chosen it.
• @mentioning an agent that has its own model runs that message on the agent's model, when your organization allows it, and the answer carries the agent's line instead of Auto's.
• Auto is saved like any other choice: a conversation on Auto stays on Auto when you reopen it, and it becomes your default for new chats.

In the Teams bot, type /model auto or choose Auto in the Settings card; each Teams answer names its model at the bottom.

Reasoning effort and thinking

Alongside the model, you control how hard it works on each message with two settings that live in the same picker — an Effort level and a Thinking toggle. Both sit in the Effort row pinned to the bottom of the model dropdown; open the dropdown and hover or click Effort to reveal them. On a phone, where there is no room beside the dropdown, Effort and Legacy models open over the model list instead, with a row at the top that takes you back to the list. From the keyboard, press the right arrow, Enter or Space on either row to open it, and the left arrow to go back (see Keyboard shortcuts).

Effort

Effort scales how thoroughly the model reasons before it answers. There are five levels:

Higher effort means more thorough responses, but each answer takes longer and uses your usage limits faster. High and above carry a warning icon so the trade-off is visible, and the current level shows on the model pill in the composer.

Your effort choice is saved to your account and applies to every new message, across all your conversations — it is not tied to a single chat the way the model or the mode is. Medium with Thinking off is the default and matches Strata's standard behavior.

Note

How each level is applied depends on the model. Some models expose a discrete reasoning-effort setting, others scale their internal thinking budget, and the newest models reason adaptively and treat the level as guidance. In every case the five levels order from lighter to deeper. If a model doesn't accept an effort setting, Strata sends your message without it — your conversation keeps working.

The Thinking toggle

The Thinking switch at the bottom of the Effort flyout turns on visible extended thinking, letting the model reason through more complex tasks step by step before it replies. It is saved to your account like the effort level, and the Effort row shows "· Thinking" next to the current level when it is on.

Some models think on every message whatever the switch says: Claude Opus 5 and 5.5, Claude Sonnet 5 and 5.5, and the Claude Fable and Mythos models. With one of them selected, the switch shows on and cannot be turned off ("Always on with …"), and Strata always shows that model's thinking; showing it costs nothing extra. Your own setting is kept for the next model you pick. GPT models that reason (GPT-5.6 and GPT-6) show a summary of their reasoning in the same place.

The reasoning trail

When a model takes a real beat to reason — more likely at higher effort or with Thinking on — Strata surfaces it:

• While the model is reasoning, a Thinking line appears with a running timer and the newest one-line summary of what it is working through.
• As soon as the model has finished thinking, the line collapses to its last summary line, before the answer or its next action appears.
• Click it to open a card titled Thought for N s, where N is the time it spent thinking, listing one line per step of its reasoning; click again to close it.

Short, simple prompts (a greeting, a quick acknowledgment) don't show a reasoning trail — it appears only when the model genuinely worked through something.

Picking a model in practice

Each model in the picker shows a short descriptor — either the model's provider name or a one-line hint from the catalog, whichever is more informative. As a starting point:

• Not sure? Choose Auto and let Strata match the model to each message (see Auto above).
• Default to a balanced model like Claude Sonnet 5.5, Claude Sonnet 5 or GPT-5.6 Terra when it is available — strong quality at good speed for everyday analysis, drafting, and data questions. Claude Sonnet 4.6, the earlier balanced default, is scheduled to retire, with Claude Sonnet 5 as its replacement; the Model Catalog shows the date.
• Switch to a top-tier model like Claude Opus 5.5, GPT-6 Astra, GPT-6 Sol, or Claude Opus 4.8 for a hard problem that needs the most reasoning — a complex SQL design, a careful financial reconciliation, a multi-step agentic build. These are slower.
• Use a fast, low-cost model (for example GPT-6 Luna, GPT-5.6 Luna, GPT-5.4 Mini, or a Grok fast tier) for quick lookups and high-volume questions where speed matters more than depth.
• Reach for DeepSeek V4 Pro to reason over very long documents at low cost. It reads text only — images and scanned pages are not sent to it.
• Cross-check a Claude answer with a GPT model (or vice versa) when you want a second opinion from a different model family.

Tip

Admins can open the catalog from Organization → Models (click Manage models, or View models without the Manage AI models permission) to see each model's capabilities, context window, knowledge cutoff, per-million-token pricing, and retirement date when one is scheduled.

How org policy shapes your list

Your administrator decides which catalog models your organization can use, and the picker only ever offers those. The available set is enforced everywhere Strata uses a chat model — chat, agents, the Teams bot, and the API. (Voice mode runs on a separate dedicated real-time model rather than the chat picker, so it's gated by whether your org has voice enabled, not by this list.)

A few things to know:

• Models your org hasn't enabled don't appear in the picker. If you need one that's missing, ask your admin to enable it in Organization → Models → Model Catalog.
• Per-user limits. An admin can narrow the list further for a single user (Organization → Users → manage a user → Access → AI models). Your effective set is the intersection of your organization's, your role's, and your personal allowed models. A built-in role doesn't restrict models, and a role bound to one of your groups adds its models to your role's — the same in the web app, Teams, scheduled agents and the API.
• Automatic fallback. If a model you previously used is later disabled for you, Strata silently switches your next message to an allowed model — your conversation keeps working. If the model was retired, Strata first switches to its replacement, then applies the same check.

Every model your organization enables is deployed in your organization's own Azure AI Foundry — that's the deployment Strata orchestrates against for each prompt. Because that inference runs inside your own Azure tenant, the prompts and data you send to these models never reach a Kronisys-controlled model. The in-product help assistant is the sole exception: Kronisys hosts it. It does not use your model catalog. See AI engine.

Legacy models

When Strata advances its lineup, older models are marked legacy and moved under Legacy models at the bottom of the picker rather than removed immediately, so you have time to move your chats, agents, and saved workflows. A model also turns legacy on its own 90 days before Microsoft retires it on Microsoft Azure AI Foundry, and the list then shows the date under its name — for example Retires Oct 19. Legacy models still run, but the AI doesn't pick them for a sub-agent.

On the retirement date the model leaves the picker. Unless your admin has turned automatic migration off, Strata moves everything that uses it to its replacement: your own model choice, your scheduled agents, and your organization's model settings. A chat that still names the retired model runs on the replacement if your organization allows it, and on another allowed model if not. When no replacement is named, or automatic migration is off, Strata doesn't change any settings, and anything that runs on its own — a scheduled agent, for example — fails until someone points it at another model. If you're on a legacy model and need migration guidance, contact your admin. See Model retirement.

Per-conversation vs. default

• The model you pick in a chat applies to that conversation. Switching mid-conversation uses the new model for your next message; messages already in the transcript are not regenerated.
• The default model for new chats is saved on your account and used until you change it. Until you pick one, you start on your organization's default model when it has one that is still current and that you're allowed to use, and otherwise on the current Claude Sonnet (Claude Sonnet 4.6 today) — in the web app and in Teams alike.
• Loading a saved conversation restores whichever model was active when you started it, so follow-up questions stay consistent. If that model has retired, Strata uses its replacement when your organization allows it; otherwise, or if the model is no longer available to you, Strata falls back to an allowed one.

Related

• Chat overview — the chat interface
• Conversation modes — auto, technical, and simple
• AI engine — the Model Catalog and bring-your-own Foundry
• Settings — set your default model