3 min read
Choose a model
Read the catalogue and choose a model for the job you actually need.
In this guide
A catalogue, not a running job
BRAND’s model catalogue draws on OpenRouter’s public model data. It can include models from OpenAI, Anthropic, Google, and xAI, subject to what the provider lists. Names, model IDs, context limits, and pricing can change; use the catalogue’s source and status when reviewing a choice.
A model selection is required to launch your brand. The worker uses that exact selection when the confirmed brand has an allocated operating budget. A catalogue listing is not a guarantee of worker compatibility: unsupported prices, fees, or output limits pause the run. The worker does not silently substitute another model.
Match the model to the task
| Question | What to look for |
|---|---|
| What must it produce? | Match the model’s supported inputs and outputs to the work. Text support alone does not establish image generation support. |
| How much context does it need? | Allow room for the brief, approved references, and the response. More context is not automatically a better result. |
| How often will it run? | Compare quality and cost over the expected workload and revisions. An uncertain paid attempt is not automatically retried. |
| Can you evaluate the result? | Use a small representative task and a clear review standard before allowing repeated paid runs. |
There is no universal best model for every brand. A detailed collection plan and a short product description may need different trade-offs. Treat provider descriptions as information to test, not a guarantee of quality.
Plan for the whole output
Image creation, image editing, mockup rendering, and print-file preparation can involve separate services and charges. A planning model should not be presented as a complete production pipeline simply because it can describe a product.
The worker reserves a conservative maximum before dispatch and settles from the provider’s actual cost. Its current scope is text only, with no web or tool calls and a 2,048-token output cap. It rejects pricing it cannot safely bound. A missing usage receipt is an unknown charge, not a free run.