When is shared model capacity different from dedicated capacity?
With shared serverless capacity, your app calls a supported model without reserving a machine. Dedicated model inference is a separate deployment option. Start by comparing the model and the workload you expect, especially whether requests arrive occasionally or keep a service busy.
Which numbers belong in the comparison?
Serverless text model rates can count input and output tokens separately. Dedicated hardware can have time-based pricing. Check the plan for the model and whether it uses shared limits, provisioned throughput or dedicated endpoints. These are different cost and availability choices.
What should you check before sending private material?
Together describes prompt storage and model-training settings under Privacy & Security. Check whether the intended service and account settings support your retention requirement. Its policy says zero retention applies after you enable it, so do not treat it as clearing past records. Keep any app-side copies under their own rules.
Would a quiet demo and a busy app need the same plan?
On paper, draw a fictional help-summary feature with ten occasional requests and another with many requests arriving together. Use the same sample text in both.
- Mark the unitsWrite the number of requests, input and output size, and when the requests arrive.
- Decide what you still need to knowList the limit and billing figures you would verify for each route. Keep the exercise fictional without launching a model or reserving capacity.