Our summary
An app often needs a small decision, not a written answer. JP Caparas explains OpenAI’s Decisions API, a service that lets software ask yes-or-no questions, choose from a fixed list or rate something on a scale. He tests it against Jev, TypeSafe’s decision model, and tries photos and long delivery logs.
You send the information and describe the allowed answers. Several independent questions can share one request, but a question cannot use another question’s answer from that request. Caparas shows how to sort support tickets, give short text enough context, offer an other option and send uncertain answers to a person.
Caparas found Jev slightly ahead on two English text tasks and cheaper when asking many questions together. OpenAI’s service handled images and much longer documents, with steadier response times in his runs. These are limited tests, not a universal ranking. Test your own examples and check how often confident answers are wrong.
Key takeaways
- Use fixed choices, yes-or-no checks or ratings when an app needs to decide what happens next. Use a writing model when it needs to compose a reply.
- Give short inputs context, include an other option and test the cutoffs that decide when uncertain answers go to a person.
- Compare cost for the whole request, including every question. Jev’s lower price mattered most in large question batches; images and long documents favoured OpenAI’s service.