Cost, multimodal, and deployment

Cost formula: input tokens × input rate + output tokens × output rate + tools/retrieval/embedding/infrastructure/review, with applicable caching and provider accounting. Prices must be checked for the real provider/model/account.

Optimization order: use code/templates where adequate → select relevant context → bound output and steps → evaluate smaller routes → safe caching → batching/traffic isolation → capacity tuning. Streaming improves visible progress; it does not automatically reduce total cost.

Multimodal: validate media → parse/OCR/vision/audio → structured observations → evidence/field checks → review → business action. Transcripts and visible damage are proposed observations, not authoritative eligibility or payment decisions.

Local/self-hosted: deployment control and possible utilization economics come with model/runtime security, GPU capacity, license, patching, and operating costs. Quantization saves memory with task-specific quality trade-offs. Fine-tune only after diagnosing a stable behavior gap and building governed training/evaluation data.