How can this be? I had great luck with GPT3 way back when… and I didn’t have function calling or chat… had to parse the JSON myself, extraction “action” and “response-text” fields… How has this been so hard for AMZN? Is it a matter of token cost and trying to use small models?

that's a reasonable theory. they've likely delayed the launch this long due to the inference cost compared to the more basic Alexa engine.

I would also guess the testing is incomplete. Alexa+ is a slow roll out so they can improve precision/recall on the intents with actual customers. Alexa+ is less deterministic than the previous model was wrt intents