Image Models
What this page shows
Section titled “What this page shows”Settings → AI models is the read-only catalog of active models in SquadOS. It brings text and image models together, lets you search by name, provider, or identifier, and shows capabilities and cost estimates.
This page does not activate models or choose which one an agent uses. Selection happens in each agent’s Generate Image with LLM tool. See Generate Image to configure the tool and validate delivery.
Access and navigation
Section titled “Access and navigation”The route is /settings/ai-models. A person needs the ai_settings.view
capability to see it; ai_settings.write does not add activation controls to
this page because the catalog is read-only.
- Open Settings → AI models.
- Select the Image filter.
- Use search to find a model by name, provider, or slug.
- Select a row to open its details on the right.
The detail view confirms the name, provider, Image type, Generates
images capability, cost estimate, and slug. An em dash (—) for cost means
the catalog does not have enough data for an estimate; it does not mean the
generation is free.
The page currently queries every globally active model and does not apply the plan catalog or exclude models marked unavailable at the provider. Therefore, Available models does not prove the organization’s entitlement or guarantee that the model will still respond. Confirm it against the plan and validate it in a test conversation before publishing.
How to read the displayed price
Section titled “How to read the displayed price”The same card changes depending on whether the organization uses its own OpenRouter key:
| Situation | UI label | What the number represents |
|---|---|---|
| Platform key | Credits/Image (est.) | estimate for one standard 1:1 generation, calculated from the model’s local catalog data |
| Own key (BYOK) | Price per image and Provider pricing | USD estimate of the cost charged directly to the organization’s OpenRouter account |
Pricing may be known per image, per megapixel, or by image-output tokens. For
token-based models, the preview uses a reference measurement stored in the
catalog. If the provider has not published pricing or the measurement is
missing, the UI displays —.
Do not copy model names or values into a permanent rule: the catalog and prices can change, and the displayed value is a preview rather than the turn’s invoice.
How actual billing works
Section titled “How actual billing works”When the agent generates an image, the backend asks OpenRouter for the call’s actual cost. That value takes precedence; the local pricing table is only a fallback when the provider does not return cost.
With the platform key, the generation, text, and other image costs from the same reply are added together and recorded in a single row linked to the assistant message. Conversion uses the total USD cost:
credits = round_up_2_decimals((total_usd_cost × 2000) + 1)The +1 applies to the accounted reply, not separately to each image.
Therefore, the Credits/Image estimate may differ from the final debit for a
reply that also generated text or more than one image.
With BYOK, OpenRouter charges the model cost directly to the organization’s
key. SquadOS records the reply in byok mode and deducts 1 operational credit
per accounted reply, even when that reply contains one or more generations.
The USD price shown in the catalog is not that operational credit.
Aspect ratio, size, and quality
Section titled “Aspect ratio, size, and quality”The native tool accepts 1:1, 16:9, 9:16, 4:3, and 3:4; 1:1 is the
default. The product sends the aspect ratio to the provider, but this tool does
not offer a 1K, 2K, or 4K resolution selector. Do not promise a specific
resolution or instruct users to increase it to 2K/4K in the current UI.
For models priced per megapixel, the backend uses an approximation close to one megapixel for each of these five ratios. The UI preview does not yet apply that multiplier and rounds to a whole credit, while the actual debit uses two decimal places. Treat the preview as an order of magnitude.
How to choose and validate
Section titled “How to choose and validate”- Compare catalog estimates and start with a lower-cost model while building the prompt.
- Confirm with the person responsible for the plan that the model belongs to the organization’s entitlement.
- In the agent, select the model under Tools → Generate Image with LLM.
- Test real requests, text inside the image, and every required aspect ratio.
- Confirm the image in the reply and monitor the balance under Settings → Subscription; with BYOK, check the OpenRouter account too.
If a model fails or disappears from the provider, choose another permitted model and test again. Do not treat a new generation as a free retry: the provider may have processed the previous attempt even when delivery failed.