Models

The AI Model is the language model that produces the agent’s replies. Your choice changes text quality, tool and attachment capabilities, speed, reasoning, and credit usage.

The catalog changes over time. Choose based on what the agent must do and validate the result in real conversations; do not rely on a static list of model names.

Open the AI Model section

Under Agents, open the agent and select Model in the editor navigation. The section shows only active text models available to the organization’s current plan. Image-generation models use a separate flow.

If the organization has no active plan, the screen says No active plan. Contact support to unlock models. If the agent already uses a model that has left the available catalog, the saved slug appears with an Unavailable badge. SquadOS preserves that selection until you choose and save another one; the badge does not switch models automatically.

The Default badge identifies the system default model. A manually created agent starts with that default — not necessarily with the Best value card.

Caution

After creating an agent, confirm that its initial model appears in the organization’s catalog. Creation currently uses the global default while the AI Model section applies the plan catalog; if the default is outside a curated plan, the agent can start as Unavailable. Choose an allowed model and save before publishing or testing.

Understand the quick picks

With no active search, up to three cards may appear:

• Cheapest — For testing;
• Best value — Recommended;
• Best performance — Maximum capability.

These labels compare models within the available catalog; they do not guarantee suitability for every task. SquadOS considers price, public capability measurements, and what the agent currently uses. A model enters these cards only if it has enough data and preserves:

• tool use when the agent has active tools;
• image input when that modality is enabled;
• a context window large enough for the current prompt;
• the tool-use level required by the number of active tools.

The SquadOS team can pin a model to one of the cards. It must still be available to the organization and meet the agent’s requirements; otherwise, the card is calculated from another candidate. The same model is not repeated across cards, so fewer than three may appear.

Recommended for this agent is a more specific marker: it appears when the analysis finds a compatible replacement for a current model below the required level. If the current model lacks enough public measurement, SquadOS does not claim that a switch is better.

Use the full list

The Search models by name, provider or slug… field filters the catalog and hides the cards while a search is active. The list is sorted by provider and name. Each row shows:

• provider and model name;
• the Default badge, where applicable;
• the measured tool-use tier: Basic, Intermediate, Powerful, Elite, or Unknown;
• output price in cr/1k, meaning credits per 1,000 output tokens.

When both kinds of candidate exist, the list is split into Suggested and Other models (not suggested for this agent). Expand the second group to inspect models that lose a capability, fall below the required level, or lack a measurement. A search opens both groups; the group also opens automatically when it contains the saved model.

You can still select a model under Other models. The choice is not blocked, but it may make the agent lose image input, context space, or reliable tool use. Reconfigure the agent or choose a suggested model before saving.

Read the details panel

The panel on the right describes the selected model:

• name, provider, and description;
• overall capability, Intelligence, and Tools, when measured by Artificial Analysis;
• Speed and Time to first word, when measured;
• Input and Output rates in credits per 1,000 tokens;
• declared capabilities such as Image, Files, Thinking, and Web Search;
• the Reasoning (extended thinking) control when the model offers it.

Unknown or not measured does not mean low capability: it means there is not enough public measurement. The current panel does not display the slug, context window, or maximum output. Use Settings → AI models for those catalog details.

Note

The speed value comes from a tokens-per-second measurement even though the current UI says “words/s.” In the EN interface, numbers in this section may also use PT-BR separators. Compare the order of magnitude and values without treating “words” as a latency guarantee.

If Capabilities is empty, do not conclude that the model accepts neither text nor tools. Models whose published modality contains text only can enter this visual state. Validate capabilities in the catalog and in an agent test.

Configure reasoning

The reasoning control appears only for models that support it. Options come from the model and may include Default, Off, Minimal, Low, Medium, High, Very high, or Maximum.

• Default lets SquadOS apply the effective default level.
• Off appears only when the provider allows reasoning to be disabled.
• For mandatory-reasoning models, the screen says Reasoning is required for this model.
• When you switch models, an unsupported level automatically returns to Default.

More reasoning can help with complex work, but it raises cost and response time. Test different levels with the same question set before adopting a higher value.

Save or discard the change

Selecting a card or row changes only the draft. The bottom bar becomes pending when the model or reasoning level changes:

1. select the model;
2. adjust reasoning if needed;
3. choose Save;
4. wait for Model saved!.

Discard restores the saved model and reasoning. Leaving the section with pending changes triggers draft protection. Test Agent uses the saved version, not a pending selection.

Understand credits and your own OpenRouter key

The Input and Output rates let you compare the variable model component. The actual debit for a message depends on submitted content, prompt, history, knowledge, attachments, tool rounds, reasoning, and response length. In SquadOS credit mode, the runtime prefers the real provider-reported cost and applies the platform formula and rounding; adding only the two screen rates does not predict the exact final debit.

With your own OpenRouter key:

• OpenRouter bills model usage directly to the linked account;
• SquadOS charges 1 fixed credit per AI message for orchestration;
• only the owner can connect, update, or remove the key.

The key is under Settings → AI provider, not AI models. See Organization Settings and Plan, Credits, and Billing.

To reduce usage, combine a suitable model with a concise prompt, enough history without excess, and a sensible response limit. Adjust history and output under Advanced Settings and image and file handling under Multimodal.

Choose and validate

Use a short, repeatable matrix:

Check	What to observe
Known answer	correctness, clarity, and prompt adherence
Missing information	admits the gap and follows the fallback rule
Required tool	chooses the right tool, sends valid arguments, and interprets the result
Unavailable tool	fails safely without inventing execution
Image or file	follows the Multimodal configuration and extracts the expected content
Long conversation	keeps required context without degrading critical instructions
Cost and latency	usage and timing fit the expected volume

Save before testing, use the same inputs to compare candidates, and verify real tool effects. Fluent text alone does not prove that the model completed the journey correctly.