How to reduce costs without killing quality? Return JSON {"tactics": [list of 2+ techniques], "why": "briefly"}.
AI calls cost money for each token—on a stream, this is the main expense item of the product. The naive scheme “we drive everything through the most powerful model” ruins. Saving levers: direct simple tasks to a cheap/fast model, and save the flagship for complex ones (routing); cache duplicate responses and prompts; limit max_tokens so that the model does not spread; cut extra context in the prompt (each extra token is money for each request); group requests (batch) where possible. Rule: 80% of requests are usually simple - sending them through the flagship is wasteful. Calculate the cost per 1000 requests before launch.